LLVM 18.0.0git
InstCombineShifts.cpp
Go to the documentation of this file.
1//===- InstCombineShifts.cpp ----------------------------------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the visitShl, visitLShr, and visitAShr functions.
10//
11//===----------------------------------------------------------------------===//
12
13#include "InstCombineInternal.h"
18using namespace llvm;
19using namespace PatternMatch;
20
21#define DEBUG_TYPE "instcombine"
22
24 Value *ShAmt1) {
25 // We have two shift amounts from two different shifts. The types of those
26 // shift amounts may not match. If that's the case let's bailout now..
27 if (ShAmt0->getType() != ShAmt1->getType())
28 return false;
29
30 // As input, we have the following pattern:
31 // Sh0 (Sh1 X, Q), K
32 // We want to rewrite that as:
33 // Sh x, (Q+K) iff (Q+K) u< bitwidth(x)
34 // While we know that originally (Q+K) would not overflow
35 // (because 2 * (N-1) u<= iN -1), we have looked past extensions of
36 // shift amounts. so it may now overflow in smaller bitwidth.
37 // To ensure that does not happen, we need to ensure that the total maximal
38 // shift amount is still representable in that smaller bit width.
39 unsigned MaximalPossibleTotalShiftAmount =
40 (Sh0->getType()->getScalarSizeInBits() - 1) +
41 (Sh1->getType()->getScalarSizeInBits() - 1);
42 APInt MaximalRepresentableShiftAmount =
44 return MaximalRepresentableShiftAmount.uge(MaximalPossibleTotalShiftAmount);
45}
46
47// Given pattern:
48// (x shiftopcode Q) shiftopcode K
49// we should rewrite it as
50// x shiftopcode (Q+K) iff (Q+K) u< bitwidth(x) and
51//
52// This is valid for any shift, but they must be identical, and we must be
53// careful in case we have (zext(Q)+zext(K)) and look past extensions,
54// (Q+K) must not overflow or else (Q+K) u< bitwidth(x) is bogus.
55//
56// AnalyzeForSignBitExtraction indicates that we will only analyze whether this
57// pattern has any 2 right-shifts that sum to 1 less than original bit width.
59 BinaryOperator *Sh0, const SimplifyQuery &SQ,
60 bool AnalyzeForSignBitExtraction) {
61 // Look for a shift of some instruction, ignore zext of shift amount if any.
62 Instruction *Sh0Op0;
63 Value *ShAmt0;
64 if (!match(Sh0,
65 m_Shift(m_Instruction(Sh0Op0), m_ZExtOrSelf(m_Value(ShAmt0)))))
66 return nullptr;
67
68 // If there is a truncation between the two shifts, we must make note of it
69 // and look through it. The truncation imposes additional constraints on the
70 // transform.
71 Instruction *Sh1;
72 Value *Trunc = nullptr;
73 match(Sh0Op0,
75 m_Instruction(Sh1)));
76
77 // Inner shift: (x shiftopcode ShAmt1)
78 // Like with other shift, ignore zext of shift amount if any.
79 Value *X, *ShAmt1;
80 if (!match(Sh1, m_Shift(m_Value(X), m_ZExtOrSelf(m_Value(ShAmt1)))))
81 return nullptr;
82
83 // Verify that it would be safe to try to add those two shift amounts.
84 if (!canTryToConstantAddTwoShiftAmounts(Sh0, ShAmt0, Sh1, ShAmt1))
85 return nullptr;
86
87 // We are only looking for signbit extraction if we have two right shifts.
88 bool HadTwoRightShifts = match(Sh0, m_Shr(m_Value(), m_Value())) &&
89 match(Sh1, m_Shr(m_Value(), m_Value()));
90 // ... and if it's not two right-shifts, we know the answer already.
91 if (AnalyzeForSignBitExtraction && !HadTwoRightShifts)
92 return nullptr;
93
94 // The shift opcodes must be identical, unless we are just checking whether
95 // this pattern can be interpreted as a sign-bit-extraction.
96 Instruction::BinaryOps ShiftOpcode = Sh0->getOpcode();
97 bool IdenticalShOpcodes = Sh0->getOpcode() == Sh1->getOpcode();
98 if (!IdenticalShOpcodes && !AnalyzeForSignBitExtraction)
99 return nullptr;
100
101 // If we saw truncation, we'll need to produce extra instruction,
102 // and for that one of the operands of the shift must be one-use,
103 // unless of course we don't actually plan to produce any instructions here.
104 if (Trunc && !AnalyzeForSignBitExtraction &&
105 !match(Sh0, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
106 return nullptr;
107
108 // Can we fold (ShAmt0+ShAmt1) ?
109 auto *NewShAmt = dyn_cast_or_null<Constant>(
110 simplifyAddInst(ShAmt0, ShAmt1, /*isNSW=*/false, /*isNUW=*/false,
111 SQ.getWithInstruction(Sh0)));
112 if (!NewShAmt)
113 return nullptr; // Did not simplify.
114 unsigned NewShAmtBitWidth = NewShAmt->getType()->getScalarSizeInBits();
115 unsigned XBitWidth = X->getType()->getScalarSizeInBits();
116 // Is the new shift amount smaller than the bit width of inner/new shift?
118 APInt(NewShAmtBitWidth, XBitWidth))))
119 return nullptr; // FIXME: could perform constant-folding.
120
121 // If there was a truncation, and we have a right-shift, we can only fold if
122 // we are left with the original sign bit. Likewise, if we were just checking
123 // that this is a sighbit extraction, this is the place to check it.
124 // FIXME: zero shift amount is also legal here, but we can't *easily* check
125 // more than one predicate so it's not really worth it.
126 if (HadTwoRightShifts && (Trunc || AnalyzeForSignBitExtraction)) {
127 // If it's not a sign bit extraction, then we're done.
128 if (!match(NewShAmt,
130 APInt(NewShAmtBitWidth, XBitWidth - 1))))
131 return nullptr;
132 // If it is, and that was the question, return the base value.
133 if (AnalyzeForSignBitExtraction)
134 return X;
135 }
136
137 assert(IdenticalShOpcodes && "Should not get here with different shifts.");
138
139 if (NewShAmt->getType() != X->getType()) {
140 NewShAmt = ConstantFoldCastOperand(Instruction::ZExt, NewShAmt,
141 X->getType(), SQ.DL);
142 if (!NewShAmt)
143 return nullptr;
144 }
145
146 // All good, we can do this fold.
147 BinaryOperator *NewShift = BinaryOperator::Create(ShiftOpcode, X, NewShAmt);
148
149 // The flags can only be propagated if there wasn't a trunc.
150 if (!Trunc) {
151 // If the pattern did not involve trunc, and both of the original shifts
152 // had the same flag set, preserve the flag.
153 if (ShiftOpcode == Instruction::BinaryOps::Shl) {
154 NewShift->setHasNoUnsignedWrap(Sh0->hasNoUnsignedWrap() &&
155 Sh1->hasNoUnsignedWrap());
156 NewShift->setHasNoSignedWrap(Sh0->hasNoSignedWrap() &&
157 Sh1->hasNoSignedWrap());
158 } else {
159 NewShift->setIsExact(Sh0->isExact() && Sh1->isExact());
160 }
161 }
162
163 Instruction *Ret = NewShift;
164 if (Trunc) {
165 Builder.Insert(NewShift);
166 Ret = CastInst::Create(Instruction::Trunc, NewShift, Sh0->getType());
167 }
168
169 return Ret;
170}
171
172// If we have some pattern that leaves only some low bits set, and then performs
173// left-shift of those bits, if none of the bits that are left after the final
174// shift are modified by the mask, we can omit the mask.
175//
176// There are many variants to this pattern:
177// a) (x & ((1 << MaskShAmt) - 1)) << ShiftShAmt
178// b) (x & (~(-1 << MaskShAmt))) << ShiftShAmt
179// c) (x & (-1 l>> MaskShAmt)) << ShiftShAmt
180// d) (x & ((-1 << MaskShAmt) l>> MaskShAmt)) << ShiftShAmt
181// e) ((x << MaskShAmt) l>> MaskShAmt) << ShiftShAmt
182// f) ((x << MaskShAmt) a>> MaskShAmt) << ShiftShAmt
183// All these patterns can be simplified to just:
184// x << ShiftShAmt
185// iff:
186// a,b) (MaskShAmt+ShiftShAmt) u>= bitwidth(x)
187// c,d,e,f) (ShiftShAmt-MaskShAmt) s>= 0 (i.e. ShiftShAmt u>= MaskShAmt)
188static Instruction *
190 const SimplifyQuery &Q,
191 InstCombiner::BuilderTy &Builder) {
192 assert(OuterShift->getOpcode() == Instruction::BinaryOps::Shl &&
193 "The input must be 'shl'!");
194
195 Value *Masked, *ShiftShAmt;
196 match(OuterShift,
197 m_Shift(m_Value(Masked), m_ZExtOrSelf(m_Value(ShiftShAmt))));
198
199 // *If* there is a truncation between an outer shift and a possibly-mask,
200 // then said truncation *must* be one-use, else we can't perform the fold.
201 Value *Trunc;
202 if (match(Masked, m_CombineAnd(m_Trunc(m_Value(Masked)), m_Value(Trunc))) &&
203 !Trunc->hasOneUse())
204 return nullptr;
205
206 Type *NarrowestTy = OuterShift->getType();
207 Type *WidestTy = Masked->getType();
208 bool HadTrunc = WidestTy != NarrowestTy;
209
210 // The mask must be computed in a type twice as wide to ensure
211 // that no bits are lost if the sum-of-shifts is wider than the base type.
212 Type *ExtendedTy = WidestTy->getExtendedType();
213
214 Value *MaskShAmt;
215
216 // ((1 << MaskShAmt) - 1)
217 auto MaskA = m_Add(m_Shl(m_One(), m_Value(MaskShAmt)), m_AllOnes());
218 // (~(-1 << maskNbits))
219 auto MaskB = m_Xor(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_AllOnes());
220 // (-1 l>> MaskShAmt)
221 auto MaskC = m_LShr(m_AllOnes(), m_Value(MaskShAmt));
222 // ((-1 << MaskShAmt) l>> MaskShAmt)
223 auto MaskD =
224 m_LShr(m_Shl(m_AllOnes(), m_Value(MaskShAmt)), m_Deferred(MaskShAmt));
225
226 Value *X;
227 Constant *NewMask;
228
229 if (match(Masked, m_c_And(m_CombineOr(MaskA, MaskB), m_Value(X)))) {
230 // Peek through an optional zext of the shift amount.
231 match(MaskShAmt, m_ZExtOrSelf(m_Value(MaskShAmt)));
232
233 // Verify that it would be safe to try to add those two shift amounts.
234 if (!canTryToConstantAddTwoShiftAmounts(OuterShift, ShiftShAmt, Masked,
235 MaskShAmt))
236 return nullptr;
237
238 // Can we simplify (MaskShAmt+ShiftShAmt) ?
239 auto *SumOfShAmts = dyn_cast_or_null<Constant>(simplifyAddInst(
240 MaskShAmt, ShiftShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q));
241 if (!SumOfShAmts)
242 return nullptr; // Did not simplify.
243 // In this pattern SumOfShAmts correlates with the number of low bits
244 // that shall remain in the root value (OuterShift).
245
246 // An extend of an undef value becomes zero because the high bits are never
247 // completely unknown. Replace the `undef` shift amounts with final
248 // shift bitwidth to ensure that the value remains undef when creating the
249 // subsequent shift op.
250 SumOfShAmts = Constant::replaceUndefsWith(
251 SumOfShAmts, ConstantInt::get(SumOfShAmts->getType()->getScalarType(),
252 ExtendedTy->getScalarSizeInBits()));
253 auto *ExtendedSumOfShAmts = ConstantFoldCastOperand(
254 Instruction::ZExt, SumOfShAmts, ExtendedTy, Q.DL);
255 if (!ExtendedSumOfShAmts)
256 return nullptr;
257
258 // And compute the mask as usual: ~(-1 << (SumOfShAmts))
259 auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(ExtendedTy);
260 auto *ExtendedInvertedMask =
261 ConstantExpr::getShl(ExtendedAllOnes, ExtendedSumOfShAmts);
262 NewMask = ConstantExpr::getNot(ExtendedInvertedMask);
263 } else if (match(Masked, m_c_And(m_CombineOr(MaskC, MaskD), m_Value(X))) ||
264 match(Masked, m_Shr(m_Shl(m_Value(X), m_Value(MaskShAmt)),
265 m_Deferred(MaskShAmt)))) {
266 // Peek through an optional zext of the shift amount.
267 match(MaskShAmt, m_ZExtOrSelf(m_Value(MaskShAmt)));
268
269 // Verify that it would be safe to try to add those two shift amounts.
270 if (!canTryToConstantAddTwoShiftAmounts(OuterShift, ShiftShAmt, Masked,
271 MaskShAmt))
272 return nullptr;
273
274 // Can we simplify (ShiftShAmt-MaskShAmt) ?
275 auto *ShAmtsDiff = dyn_cast_or_null<Constant>(simplifySubInst(
276 ShiftShAmt, MaskShAmt, /*IsNSW=*/false, /*IsNUW=*/false, Q));
277 if (!ShAmtsDiff)
278 return nullptr; // Did not simplify.
279 // In this pattern ShAmtsDiff correlates with the number of high bits that
280 // shall be unset in the root value (OuterShift).
281
282 // An extend of an undef value becomes zero because the high bits are never
283 // completely unknown. Replace the `undef` shift amounts with negated
284 // bitwidth of innermost shift to ensure that the value remains undef when
285 // creating the subsequent shift op.
286 unsigned WidestTyBitWidth = WidestTy->getScalarSizeInBits();
287 ShAmtsDiff = Constant::replaceUndefsWith(
288 ShAmtsDiff, ConstantInt::get(ShAmtsDiff->getType()->getScalarType(),
289 -WidestTyBitWidth));
290 auto *ExtendedNumHighBitsToClear = ConstantFoldCastOperand(
291 Instruction::ZExt,
292 ConstantExpr::getSub(ConstantInt::get(ShAmtsDiff->getType(),
293 WidestTyBitWidth,
294 /*isSigned=*/false),
295 ShAmtsDiff),
296 ExtendedTy, Q.DL);
297 if (!ExtendedNumHighBitsToClear)
298 return nullptr;
299
300 // And compute the mask as usual: (-1 l>> (NumHighBitsToClear))
301 auto *ExtendedAllOnes = ConstantExpr::getAllOnesValue(ExtendedTy);
302 NewMask = ConstantFoldBinaryOpOperands(Instruction::LShr, ExtendedAllOnes,
303 ExtendedNumHighBitsToClear, Q.DL);
304 if (!NewMask)
305 return nullptr;
306 } else
307 return nullptr; // Don't know anything about this pattern.
308
309 NewMask = ConstantExpr::getTrunc(NewMask, NarrowestTy);
310
311 // Does this mask has any unset bits? If not then we can just not apply it.
312 bool NeedMask = !match(NewMask, m_AllOnes());
313
314 // If we need to apply a mask, there are several more restrictions we have.
315 if (NeedMask) {
316 // The old masking instruction must go away.
317 if (!Masked->hasOneUse())
318 return nullptr;
319 // The original "masking" instruction must not have been`ashr`.
320 if (match(Masked, m_AShr(m_Value(), m_Value())))
321 return nullptr;
322 }
323
324 // If we need to apply truncation, let's do it first, since we can.
325 // We have already ensured that the old truncation will go away.
326 if (HadTrunc)
327 X = Builder.CreateTrunc(X, NarrowestTy);
328
329 // No 'NUW'/'NSW'! We no longer know that we won't shift-out non-0 bits.
330 // We didn't change the Type of this outermost shift, so we can just do it.
331 auto *NewShift = BinaryOperator::Create(OuterShift->getOpcode(), X,
332 OuterShift->getOperand(1));
333 if (!NeedMask)
334 return NewShift;
335
336 Builder.Insert(NewShift);
337 return BinaryOperator::Create(Instruction::And, NewShift, NewMask);
338}
339
340/// If we have a shift-by-constant of a bin op (bitwise logic op or add/sub w/
341/// shl) that itself has a shift-by-constant operand with identical opcode, we
342/// may be able to convert that into 2 independent shifts followed by the logic
343/// op. This eliminates a use of an intermediate value (reduces dependency
344/// chain).
346 InstCombiner::BuilderTy &Builder) {
347 assert(I.isShift() && "Expected a shift as input");
348 auto *BinInst = dyn_cast<BinaryOperator>(I.getOperand(0));
349 if (!BinInst ||
350 (!BinInst->isBitwiseLogicOp() &&
351 BinInst->getOpcode() != Instruction::Add &&
352 BinInst->getOpcode() != Instruction::Sub) ||
353 !BinInst->hasOneUse())
354 return nullptr;
355
356 Constant *C0, *C1;
357 if (!match(I.getOperand(1), m_Constant(C1)))
358 return nullptr;
359
360 Instruction::BinaryOps ShiftOpcode = I.getOpcode();
361 // Transform for add/sub only works with shl.
362 if ((BinInst->getOpcode() == Instruction::Add ||
363 BinInst->getOpcode() == Instruction::Sub) &&
364 ShiftOpcode != Instruction::Shl)
365 return nullptr;
366
367 Type *Ty = I.getType();
368
369 // Find a matching one-use shift by constant. The fold is not valid if the sum
370 // of the shift values equals or exceeds bitwidth.
371 // TODO: Remove the one-use check if the other logic operand (Y) is constant.
372 Value *X, *Y;
373 auto matchFirstShift = [&](Value *V) {
374 APInt Threshold(Ty->getScalarSizeInBits(), Ty->getScalarSizeInBits());
375 return match(V,
376 m_OneUse(m_BinOp(ShiftOpcode, m_Value(X), m_Constant(C0)))) &&
379 };
380
381 // Logic ops and Add are commutative, so check each operand for a match. Sub
382 // is not so we cannot reoder if we match operand(1) and need to keep the
383 // operands in their original positions.
384 bool FirstShiftIsOp1 = false;
385 if (matchFirstShift(BinInst->getOperand(0)))
386 Y = BinInst->getOperand(1);
387 else if (matchFirstShift(BinInst->getOperand(1))) {
388 Y = BinInst->getOperand(0);
389 FirstShiftIsOp1 = BinInst->getOpcode() == Instruction::Sub;
390 } else
391 return nullptr;
392
393 // shift (binop (shift X, C0), Y), C1 -> binop (shift X, C0+C1), (shift Y, C1)
394 Constant *ShiftSumC = ConstantExpr::getAdd(C0, C1);
395 Value *NewShift1 = Builder.CreateBinOp(ShiftOpcode, X, ShiftSumC);
396 Value *NewShift2 = Builder.CreateBinOp(ShiftOpcode, Y, C1);
397 Value *Op1 = FirstShiftIsOp1 ? NewShift2 : NewShift1;
398 Value *Op2 = FirstShiftIsOp1 ? NewShift1 : NewShift2;
399 return BinaryOperator::Create(BinInst->getOpcode(), Op1, Op2);
400}
401
404 return Phi;
405
406 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
407 assert(Op0->getType() == Op1->getType());
408 Type *Ty = I.getType();
409
410 // If the shift amount is a one-use `sext`, we can demote it to `zext`.
411 Value *Y;
412 if (match(Op1, m_OneUse(m_SExt(m_Value(Y))))) {
413 Value *NewExt = Builder.CreateZExt(Y, Ty, Op1->getName());
414 return BinaryOperator::Create(I.getOpcode(), Op0, NewExt);
415 }
416
417 // See if we can fold away this shift.
419 return &I;
420
421 // Try to fold constant and into select arguments.
422 if (isa<Constant>(Op0))
423 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
424 if (Instruction *R = FoldOpIntoSelect(I, SI))
425 return R;
426
427 if (Constant *CUI = dyn_cast<Constant>(Op1))
428 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
429 return Res;
430
431 if (auto *NewShift = cast_or_null<Instruction>(
433 return NewShift;
434
435 // Pre-shift a constant shifted by a variable amount with constant offset:
436 // C shift (A add nuw C1) --> (C shift C1) shift A
437 Value *A;
438 Constant *C, *C1;
439 if (match(Op0, m_Constant(C)) &&
440 match(Op1, m_NUWAdd(m_Value(A), m_Constant(C1)))) {
441 Value *NewC = Builder.CreateBinOp(I.getOpcode(), C, C1);
442 return BinaryOperator::Create(I.getOpcode(), NewC, A);
443 }
444
445 unsigned BitWidth = Ty->getScalarSizeInBits();
446
447 const APInt *AC, *AddC;
448 // Try to pre-shift a constant shifted by a variable amount added with a
449 // negative number:
450 // C << (X - AddC) --> (C >> AddC) << X
451 // and
452 // C >> (X - AddC) --> (C << AddC) >> X
453 if (match(Op0, m_APInt(AC)) && match(Op1, m_Add(m_Value(A), m_APInt(AddC))) &&
454 AddC->isNegative() && (-*AddC).ult(BitWidth)) {
455 assert(!AC->isZero() && "Expected simplify of shifted zero");
456 unsigned PosOffset = (-*AddC).getZExtValue();
457
458 auto isSuitableForPreShift = [PosOffset, &I, AC]() {
459 switch (I.getOpcode()) {
460 default:
461 return false;
462 case Instruction::Shl:
463 return (I.hasNoSignedWrap() || I.hasNoUnsignedWrap()) &&
464 AC->eq(AC->lshr(PosOffset).shl(PosOffset));
465 case Instruction::LShr:
466 return I.isExact() && AC->eq(AC->shl(PosOffset).lshr(PosOffset));
467 case Instruction::AShr:
468 return I.isExact() && AC->eq(AC->shl(PosOffset).ashr(PosOffset));
469 }
470 };
471 if (isSuitableForPreShift()) {
472 Constant *NewC = ConstantInt::get(Ty, I.getOpcode() == Instruction::Shl
473 ? AC->lshr(PosOffset)
474 : AC->shl(PosOffset));
475 BinaryOperator *NewShiftOp =
476 BinaryOperator::Create(I.getOpcode(), NewC, A);
477 if (I.getOpcode() == Instruction::Shl) {
478 NewShiftOp->setHasNoUnsignedWrap(I.hasNoUnsignedWrap());
479 } else {
480 NewShiftOp->setIsExact();
481 }
482 return NewShiftOp;
483 }
484 }
485
486 // X shift (A srem C) -> X shift (A and (C - 1)) iff C is a power of 2.
487 // Because shifts by negative values (which could occur if A were negative)
488 // are undefined.
489 if (Op1->hasOneUse() && match(Op1, m_SRem(m_Value(A), m_Constant(C))) &&
490 match(C, m_Power2())) {
491 // FIXME: Should this get moved into SimplifyDemandedBits by saying we don't
492 // demand the sign bit (and many others) here??
494 Value *Rem = Builder.CreateAnd(A, Mask, Op1->getName());
495 return replaceOperand(I, 1, Rem);
496 }
497
499 return Logic;
500
501 if (match(Op1, m_Or(m_Value(), m_SpecificInt(BitWidth - 1))))
502 return replaceOperand(I, 1, ConstantInt::get(Ty, BitWidth - 1));
503
504 return nullptr;
505}
506
507/// Return true if we can simplify two logical (either left or right) shifts
508/// that have constant shift amounts: OuterShift (InnerShift X, C1), C2.
509static bool canEvaluateShiftedShift(unsigned OuterShAmt, bool IsOuterShl,
510 Instruction *InnerShift,
511 InstCombinerImpl &IC, Instruction *CxtI) {
512 assert(InnerShift->isLogicalShift() && "Unexpected instruction type");
513
514 // We need constant scalar or constant splat shifts.
515 const APInt *InnerShiftConst;
516 if (!match(InnerShift->getOperand(1), m_APInt(InnerShiftConst)))
517 return false;
518
519 // Two logical shifts in the same direction:
520 // shl (shl X, C1), C2 --> shl X, C1 + C2
521 // lshr (lshr X, C1), C2 --> lshr X, C1 + C2
522 bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl;
523 if (IsInnerShl == IsOuterShl)
524 return true;
525
526 // Equal shift amounts in opposite directions become bitwise 'and':
527 // lshr (shl X, C), C --> and X, C'
528 // shl (lshr X, C), C --> and X, C'
529 if (*InnerShiftConst == OuterShAmt)
530 return true;
531
532 // If the 2nd shift is bigger than the 1st, we can fold:
533 // lshr (shl X, C1), C2 --> and (shl X, C1 - C2), C3
534 // shl (lshr X, C1), C2 --> and (lshr X, C1 - C2), C3
535 // but it isn't profitable unless we know the and'd out bits are already zero.
536 // Also, check that the inner shift is valid (less than the type width) or
537 // we'll crash trying to produce the bit mask for the 'and'.
538 unsigned TypeWidth = InnerShift->getType()->getScalarSizeInBits();
539 if (InnerShiftConst->ugt(OuterShAmt) && InnerShiftConst->ult(TypeWidth)) {
540 unsigned InnerShAmt = InnerShiftConst->getZExtValue();
541 unsigned MaskShift =
542 IsInnerShl ? TypeWidth - InnerShAmt : InnerShAmt - OuterShAmt;
543 APInt Mask = APInt::getLowBitsSet(TypeWidth, OuterShAmt) << MaskShift;
544 if (IC.MaskedValueIsZero(InnerShift->getOperand(0), Mask, 0, CxtI))
545 return true;
546 }
547
548 return false;
549}
550
551/// See if we can compute the specified value, but shifted logically to the left
552/// or right by some number of bits. This should return true if the expression
553/// can be computed for the same cost as the current expression tree. This is
554/// used to eliminate extraneous shifting from things like:
555/// %C = shl i128 %A, 64
556/// %D = shl i128 %B, 96
557/// %E = or i128 %C, %D
558/// %F = lshr i128 %E, 64
559/// where the client will ask if E can be computed shifted right by 64-bits. If
560/// this succeeds, getShiftedValue() will be called to produce the value.
561static bool canEvaluateShifted(Value *V, unsigned NumBits, bool IsLeftShift,
562 InstCombinerImpl &IC, Instruction *CxtI) {
563 // We can always evaluate immediate constants.
564 if (match(V, m_ImmConstant()))
565 return true;
566
567 Instruction *I = dyn_cast<Instruction>(V);
568 if (!I) return false;
569
570 // We can't mutate something that has multiple uses: doing so would
571 // require duplicating the instruction in general, which isn't profitable.
572 if (!I->hasOneUse()) return false;
573
574 switch (I->getOpcode()) {
575 default: return false;
576 case Instruction::And:
577 case Instruction::Or:
578 case Instruction::Xor:
579 // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
580 return canEvaluateShifted(I->getOperand(0), NumBits, IsLeftShift, IC, I) &&
581 canEvaluateShifted(I->getOperand(1), NumBits, IsLeftShift, IC, I);
582
583 case Instruction::Shl:
584 case Instruction::LShr:
585 return canEvaluateShiftedShift(NumBits, IsLeftShift, I, IC, CxtI);
586
587 case Instruction::Select: {
588 SelectInst *SI = cast<SelectInst>(I);
589 Value *TrueVal = SI->getTrueValue();
590 Value *FalseVal = SI->getFalseValue();
591 return canEvaluateShifted(TrueVal, NumBits, IsLeftShift, IC, SI) &&
592 canEvaluateShifted(FalseVal, NumBits, IsLeftShift, IC, SI);
593 }
594 case Instruction::PHI: {
595 // We can change a phi if we can change all operands. Note that we never
596 // get into trouble with cyclic PHIs here because we only consider
597 // instructions with a single use.
598 PHINode *PN = cast<PHINode>(I);
599 for (Value *IncValue : PN->incoming_values())
600 if (!canEvaluateShifted(IncValue, NumBits, IsLeftShift, IC, PN))
601 return false;
602 return true;
603 }
604 case Instruction::Mul: {
605 const APInt *MulConst;
606 // We can fold (shr (mul X, -(1 << C)), C) -> (and (neg X), C`)
607 return !IsLeftShift && match(I->getOperand(1), m_APInt(MulConst)) &&
608 MulConst->isNegatedPowerOf2() && MulConst->countr_zero() == NumBits;
609 }
610 }
611}
612
613/// Fold OuterShift (InnerShift X, C1), C2.
614/// See canEvaluateShiftedShift() for the constraints on these instructions.
615static Value *foldShiftedShift(BinaryOperator *InnerShift, unsigned OuterShAmt,
616 bool IsOuterShl,
617 InstCombiner::BuilderTy &Builder) {
618 bool IsInnerShl = InnerShift->getOpcode() == Instruction::Shl;
619 Type *ShType = InnerShift->getType();
620 unsigned TypeWidth = ShType->getScalarSizeInBits();
621
622 // We only accept shifts-by-a-constant in canEvaluateShifted().
623 const APInt *C1;
624 match(InnerShift->getOperand(1), m_APInt(C1));
625 unsigned InnerShAmt = C1->getZExtValue();
626
627 // Change the shift amount and clear the appropriate IR flags.
628 auto NewInnerShift = [&](unsigned ShAmt) {
629 InnerShift->setOperand(1, ConstantInt::get(ShType, ShAmt));
630 if (IsInnerShl) {
631 InnerShift->setHasNoUnsignedWrap(false);
632 InnerShift->setHasNoSignedWrap(false);
633 } else {
634 InnerShift->setIsExact(false);
635 }
636 return InnerShift;
637 };
638
639 // Two logical shifts in the same direction:
640 // shl (shl X, C1), C2 --> shl X, C1 + C2
641 // lshr (lshr X, C1), C2 --> lshr X, C1 + C2
642 if (IsInnerShl == IsOuterShl) {
643 // If this is an oversized composite shift, then unsigned shifts get 0.
644 if (InnerShAmt + OuterShAmt >= TypeWidth)
645 return Constant::getNullValue(ShType);
646
647 return NewInnerShift(InnerShAmt + OuterShAmt);
648 }
649
650 // Equal shift amounts in opposite directions become bitwise 'and':
651 // lshr (shl X, C), C --> and X, C'
652 // shl (lshr X, C), C --> and X, C'
653 if (InnerShAmt == OuterShAmt) {
654 APInt Mask = IsInnerShl
655 ? APInt::getLowBitsSet(TypeWidth, TypeWidth - OuterShAmt)
656 : APInt::getHighBitsSet(TypeWidth, TypeWidth - OuterShAmt);
657 Value *And = Builder.CreateAnd(InnerShift->getOperand(0),
658 ConstantInt::get(ShType, Mask));
659 if (auto *AndI = dyn_cast<Instruction>(And)) {
660 AndI->moveBefore(InnerShift);
661 AndI->takeName(InnerShift);
662 }
663 return And;
664 }
665
666 assert(InnerShAmt > OuterShAmt &&
667 "Unexpected opposite direction logical shift pair");
668
669 // In general, we would need an 'and' for this transform, but
670 // canEvaluateShiftedShift() guarantees that the masked-off bits are not used.
671 // lshr (shl X, C1), C2 --> shl X, C1 - C2
672 // shl (lshr X, C1), C2 --> lshr X, C1 - C2
673 return NewInnerShift(InnerShAmt - OuterShAmt);
674}
675
676/// When canEvaluateShifted() returns true for an expression, this function
677/// inserts the new computation that produces the shifted value.
678static Value *getShiftedValue(Value *V, unsigned NumBits, bool isLeftShift,
679 InstCombinerImpl &IC, const DataLayout &DL) {
680 // We can always evaluate constants shifted.
681 if (Constant *C = dyn_cast<Constant>(V)) {
682 if (isLeftShift)
683 return IC.Builder.CreateShl(C, NumBits);
684 else
685 return IC.Builder.CreateLShr(C, NumBits);
686 }
687
688 Instruction *I = cast<Instruction>(V);
689 IC.addToWorklist(I);
690
691 switch (I->getOpcode()) {
692 default: llvm_unreachable("Inconsistency with CanEvaluateShifted");
693 case Instruction::And:
694 case Instruction::Or:
695 case Instruction::Xor:
696 // Bitwise operators can all arbitrarily be arbitrarily evaluated shifted.
697 I->setOperand(
698 0, getShiftedValue(I->getOperand(0), NumBits, isLeftShift, IC, DL));
699 I->setOperand(
700 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL));
701 return I;
702
703 case Instruction::Shl:
704 case Instruction::LShr:
705 return foldShiftedShift(cast<BinaryOperator>(I), NumBits, isLeftShift,
706 IC.Builder);
707
708 case Instruction::Select:
709 I->setOperand(
710 1, getShiftedValue(I->getOperand(1), NumBits, isLeftShift, IC, DL));
711 I->setOperand(
712 2, getShiftedValue(I->getOperand(2), NumBits, isLeftShift, IC, DL));
713 return I;
714 case Instruction::PHI: {
715 // We can change a phi if we can change all operands. Note that we never
716 // get into trouble with cyclic PHIs here because we only consider
717 // instructions with a single use.
718 PHINode *PN = cast<PHINode>(I);
719 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
721 isLeftShift, IC, DL));
722 return PN;
723 }
724 case Instruction::Mul: {
725 assert(!isLeftShift && "Unexpected shift direction!");
726 auto *Neg = BinaryOperator::CreateNeg(I->getOperand(0));
727 IC.InsertNewInstWith(Neg, I->getIterator());
728 unsigned TypeWidth = I->getType()->getScalarSizeInBits();
729 APInt Mask = APInt::getLowBitsSet(TypeWidth, TypeWidth - NumBits);
730 auto *And = BinaryOperator::CreateAnd(Neg,
731 ConstantInt::get(I->getType(), Mask));
732 And->takeName(I);
733 return IC.InsertNewInstWith(And, I->getIterator());
734 }
735 }
736}
737
738// If this is a bitwise operator or add with a constant RHS we might be able
739// to pull it through a shift.
741 BinaryOperator *BO) {
742 switch (BO->getOpcode()) {
743 default:
744 return false; // Do not perform transform!
745 case Instruction::Add:
746 return Shift.getOpcode() == Instruction::Shl;
747 case Instruction::Or:
748 case Instruction::And:
749 return true;
750 case Instruction::Xor:
751 // Do not change a 'not' of logical shift because that would create a normal
752 // 'xor'. The 'not' is likely better for analysis, SCEV, and codegen.
753 return !(Shift.isLogicalShift() && match(BO, m_Not(m_Value())));
754 }
755}
756
758 BinaryOperator &I) {
759 // (C2 << X) << C1 --> (C2 << C1) << X
760 // (C2 >> X) >> C1 --> (C2 >> C1) >> X
761 Constant *C2;
762 Value *X;
763 if (match(Op0, m_BinOp(I.getOpcode(), m_ImmConstant(C2), m_Value(X))))
765 I.getOpcode(), Builder.CreateBinOp(I.getOpcode(), C2, C1), X);
766
767 bool IsLeftShift = I.getOpcode() == Instruction::Shl;
768 Type *Ty = I.getType();
769 unsigned TypeBits = Ty->getScalarSizeInBits();
770
771 // (X / +DivC) >> (Width - 1) --> ext (X <= -DivC)
772 // (X / -DivC) >> (Width - 1) --> ext (X >= +DivC)
773 const APInt *DivC;
774 if (!IsLeftShift && match(C1, m_SpecificIntAllowUndef(TypeBits - 1)) &&
775 match(Op0, m_SDiv(m_Value(X), m_APInt(DivC))) && !DivC->isZero() &&
776 !DivC->isMinSignedValue()) {
777 Constant *NegDivC = ConstantInt::get(Ty, -(*DivC));
780 Value *Cmp = Builder.CreateICmp(Pred, X, NegDivC);
781 auto ExtOpcode = (I.getOpcode() == Instruction::AShr) ? Instruction::SExt
782 : Instruction::ZExt;
783 return CastInst::Create(ExtOpcode, Cmp, Ty);
784 }
785
786 const APInt *Op1C;
787 if (!match(C1, m_APInt(Op1C)))
788 return nullptr;
789
790 assert(!Op1C->uge(TypeBits) &&
791 "Shift over the type width should have been removed already");
792
793 // See if we can propagate this shift into the input, this covers the trivial
794 // cast of lshr(shl(x,c1),c2) as well as other more complex cases.
795 if (I.getOpcode() != Instruction::AShr &&
796 canEvaluateShifted(Op0, Op1C->getZExtValue(), IsLeftShift, *this, &I)) {
798 dbgs() << "ICE: GetShiftedValue propagating shift through expression"
799 " to eliminate shift:\n IN: "
800 << *Op0 << "\n SH: " << I << "\n");
801
802 return replaceInstUsesWith(
803 I, getShiftedValue(Op0, Op1C->getZExtValue(), IsLeftShift, *this, DL));
804 }
805
806 if (Instruction *FoldedShift = foldBinOpIntoSelectOrPhi(I))
807 return FoldedShift;
808
809 if (!Op0->hasOneUse())
810 return nullptr;
811
812 if (auto *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
813 // If the operand is a bitwise operator with a constant RHS, and the
814 // shift is the only use, we can pull it out of the shift.
815 const APInt *Op0C;
816 if (match(Op0BO->getOperand(1), m_APInt(Op0C))) {
817 if (canShiftBinOpWithConstantRHS(I, Op0BO)) {
818 Value *NewRHS =
819 Builder.CreateBinOp(I.getOpcode(), Op0BO->getOperand(1), C1);
820
821 Value *NewShift =
822 Builder.CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), C1);
823 NewShift->takeName(Op0BO);
824
825 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift, NewRHS);
826 }
827 }
828 }
829
830 // If we have a select that conditionally executes some binary operator,
831 // see if we can pull it the select and operator through the shift.
832 //
833 // For example, turning:
834 // shl (select C, (add X, C1), X), C2
835 // Into:
836 // Y = shl X, C2
837 // select C, (add Y, C1 << C2), Y
838 Value *Cond;
839 BinaryOperator *TBO;
840 Value *FalseVal;
841 if (match(Op0, m_Select(m_Value(Cond), m_OneUse(m_BinOp(TBO)),
842 m_Value(FalseVal)))) {
843 const APInt *C;
844 if (!isa<Constant>(FalseVal) && TBO->getOperand(0) == FalseVal &&
845 match(TBO->getOperand(1), m_APInt(C)) &&
847 Value *NewRHS =
848 Builder.CreateBinOp(I.getOpcode(), TBO->getOperand(1), C1);
849
850 Value *NewShift = Builder.CreateBinOp(I.getOpcode(), FalseVal, C1);
851 Value *NewOp = Builder.CreateBinOp(TBO->getOpcode(), NewShift, NewRHS);
852 return SelectInst::Create(Cond, NewOp, NewShift);
853 }
854 }
855
856 BinaryOperator *FBO;
857 Value *TrueVal;
858 if (match(Op0, m_Select(m_Value(Cond), m_Value(TrueVal),
859 m_OneUse(m_BinOp(FBO))))) {
860 const APInt *C;
861 if (!isa<Constant>(TrueVal) && FBO->getOperand(0) == TrueVal &&
862 match(FBO->getOperand(1), m_APInt(C)) &&
864 Value *NewRHS =
865 Builder.CreateBinOp(I.getOpcode(), FBO->getOperand(1), C1);
866
867 Value *NewShift = Builder.CreateBinOp(I.getOpcode(), TrueVal, C1);
868 Value *NewOp = Builder.CreateBinOp(FBO->getOpcode(), NewShift, NewRHS);
869 return SelectInst::Create(Cond, NewShift, NewOp);
870 }
871 }
872
873 return nullptr;
874}
875
876// Tries to perform
877// (lshr (add (zext X), (zext Y)), K)
878// -> (icmp ult (add X, Y), X)
879// where
880// - The add's operands are zexts from a K-bits integer to a bigger type.
881// - The add is only used by the shr, or by iK (or narrower) truncates.
882// - The lshr type has more than 2 bits (other types are boolean math).
883// - K > 1
884// note that
885// - The resulting add cannot have nuw/nsw, else on overflow we get a
886// poison value and the transform isn't legal anymore.
887Instruction *InstCombinerImpl::foldLShrOverflowBit(BinaryOperator &I) {
888 assert(I.getOpcode() == Instruction::LShr);
889
890 Value *Add = I.getOperand(0);
891 Value *ShiftAmt = I.getOperand(1);
892 Type *Ty = I.getType();
893
894 if (Ty->getScalarSizeInBits() < 3)
895 return nullptr;
896
897 const APInt *ShAmtAPInt = nullptr;
898 Value *X = nullptr, *Y = nullptr;
899 if (!match(ShiftAmt, m_APInt(ShAmtAPInt)) ||
900 !match(Add,
902 return nullptr;
903
904 const unsigned ShAmt = ShAmtAPInt->getZExtValue();
905 if (ShAmt == 1)
906 return nullptr;
907
908 // X/Y are zexts from `ShAmt`-sized ints.
909 if (X->getType()->getScalarSizeInBits() != ShAmt ||
910 Y->getType()->getScalarSizeInBits() != ShAmt)
911 return nullptr;
912
913 // Make sure that `Add` is only used by `I` and `ShAmt`-truncates.
914 if (!Add->hasOneUse()) {
915 for (User *U : Add->users()) {
916 if (U == &I)
917 continue;
918
919 TruncInst *Trunc = dyn_cast<TruncInst>(U);
920 if (!Trunc || Trunc->getType()->getScalarSizeInBits() > ShAmt)
921 return nullptr;
922 }
923 }
924
925 // Insert at Add so that the newly created `NarrowAdd` will dominate it's
926 // users (i.e. `Add`'s users).
927 Instruction *AddInst = cast<Instruction>(Add);
928 Builder.SetInsertPoint(AddInst);
929
930 Value *NarrowAdd = Builder.CreateAdd(X, Y, "add.narrowed");
931 Value *Overflow =
932 Builder.CreateICmpULT(NarrowAdd, X, "add.narrowed.overflow");
933
934 // Replace the uses of the original add with a zext of the
935 // NarrowAdd's result. Note that all users at this stage are known to
936 // be ShAmt-sized truncs, or the lshr itself.
937 if (!Add->hasOneUse()) {
938 replaceInstUsesWith(*AddInst, Builder.CreateZExt(NarrowAdd, Ty));
939 eraseInstFromFunction(*AddInst);
940 }
941
942 // Replace the LShr with a zext of the overflow check.
943 return new ZExtInst(Overflow, Ty);
944}
945
946// Try to set nuw/nsw flags on shl or exact flag on lshr/ashr using knownbits.
948 assert(I.isShift() && "Expected a shift as input");
949 // We already have all the flags.
950 if (I.getOpcode() == Instruction::Shl) {
951 if (I.hasNoUnsignedWrap() && I.hasNoSignedWrap())
952 return false;
953 } else {
954 if (I.isExact())
955 return false;
956
957 // shr (shl X, Y), Y
958 if (match(I.getOperand(0), m_Shl(m_Value(), m_Specific(I.getOperand(1))))) {
959 I.setIsExact();
960 return true;
961 }
962 }
963
964 // Compute what we know about shift count.
965 KnownBits KnownCnt = computeKnownBits(I.getOperand(1), /* Depth */ 0, Q);
966 unsigned BitWidth = KnownCnt.getBitWidth();
967 // Since shift produces a poison value if RHS is equal to or larger than the
968 // bit width, we can safely assume that RHS is less than the bit width.
969 uint64_t MaxCnt = KnownCnt.getMaxValue().getLimitedValue(BitWidth - 1);
970
971 KnownBits KnownAmt = computeKnownBits(I.getOperand(0), /* Depth */ 0, Q);
972 bool Changed = false;
973
974 if (I.getOpcode() == Instruction::Shl) {
975 // If we have as many leading zeros than maximum shift cnt we have nuw.
976 if (!I.hasNoUnsignedWrap() && MaxCnt <= KnownAmt.countMinLeadingZeros()) {
977 I.setHasNoUnsignedWrap();
978 Changed = true;
979 }
980 // If we have more sign bits than maximum shift cnt we have nsw.
981 if (!I.hasNoSignedWrap()) {
982 if (MaxCnt < KnownAmt.countMinSignBits() ||
983 MaxCnt < ComputeNumSignBits(I.getOperand(0), Q.DL, /*Depth*/ 0, Q.AC,
984 Q.CxtI, Q.DT)) {
985 I.setHasNoSignedWrap();
986 Changed = true;
987 }
988 }
989 return Changed;
990 }
991
992 // If we have at least as many trailing zeros as maximum count then we have
993 // exact.
994 Changed = MaxCnt <= KnownAmt.countMinTrailingZeros();
995 I.setIsExact(Changed);
996
997 return Changed;
998}
999
1002
1003 if (Value *V = simplifyShlInst(I.getOperand(0), I.getOperand(1),
1004 I.hasNoSignedWrap(), I.hasNoUnsignedWrap(), Q))
1005 return replaceInstUsesWith(I, V);
1006
1008 return X;
1009
1011 return V;
1012
1014 return V;
1015
1016 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1017 Type *Ty = I.getType();
1018 unsigned BitWidth = Ty->getScalarSizeInBits();
1019
1020 const APInt *C;
1021 if (match(Op1, m_APInt(C))) {
1022 unsigned ShAmtC = C->getZExtValue();
1023
1024 // shl (zext X), C --> zext (shl X, C)
1025 // This is only valid if X would have zeros shifted out.
1026 Value *X;
1027 if (match(Op0, m_OneUse(m_ZExt(m_Value(X))))) {
1028 unsigned SrcWidth = X->getType()->getScalarSizeInBits();
1029 if (ShAmtC < SrcWidth &&
1030 MaskedValueIsZero(X, APInt::getHighBitsSet(SrcWidth, ShAmtC), 0, &I))
1031 return new ZExtInst(Builder.CreateShl(X, ShAmtC), Ty);
1032 }
1033
1034 // (X >> C) << C --> X & (-1 << C)
1035 if (match(Op0, m_Shr(m_Value(X), m_Specific(Op1)))) {
1037 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask));
1038 }
1039
1040 const APInt *C1;
1041 if (match(Op0, m_Exact(m_Shr(m_Value(X), m_APInt(C1)))) &&
1042 C1->ult(BitWidth)) {
1043 unsigned ShrAmt = C1->getZExtValue();
1044 if (ShrAmt < ShAmtC) {
1045 // If C1 < C: (X >>?,exact C1) << C --> X << (C - C1)
1046 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmtC - ShrAmt);
1047 auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
1048 NewShl->setHasNoUnsignedWrap(
1049 I.hasNoUnsignedWrap() ||
1050 (ShrAmt &&
1051 cast<Instruction>(Op0)->getOpcode() == Instruction::LShr &&
1052 I.hasNoSignedWrap()));
1053 NewShl->setHasNoSignedWrap(I.hasNoSignedWrap());
1054 return NewShl;
1055 }
1056 if (ShrAmt > ShAmtC) {
1057 // If C1 > C: (X >>?exact C1) << C --> X >>?exact (C1 - C)
1058 Constant *ShiftDiff = ConstantInt::get(Ty, ShrAmt - ShAmtC);
1059 auto *NewShr = BinaryOperator::Create(
1060 cast<BinaryOperator>(Op0)->getOpcode(), X, ShiftDiff);
1061 NewShr->setIsExact(true);
1062 return NewShr;
1063 }
1064 }
1065
1066 if (match(Op0, m_OneUse(m_Shr(m_Value(X), m_APInt(C1)))) &&
1067 C1->ult(BitWidth)) {
1068 unsigned ShrAmt = C1->getZExtValue();
1069 if (ShrAmt < ShAmtC) {
1070 // If C1 < C: (X >>? C1) << C --> (X << (C - C1)) & (-1 << C)
1071 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmtC - ShrAmt);
1072 auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
1073 NewShl->setHasNoUnsignedWrap(
1074 I.hasNoUnsignedWrap() ||
1075 (ShrAmt &&
1076 cast<Instruction>(Op0)->getOpcode() == Instruction::LShr &&
1077 I.hasNoSignedWrap()));
1078 NewShl->setHasNoSignedWrap(I.hasNoSignedWrap());
1079 Builder.Insert(NewShl);
1081 return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask));
1082 }
1083 if (ShrAmt > ShAmtC) {
1084 // If C1 > C: (X >>? C1) << C --> (X >>? (C1 - C)) & (-1 << C)
1085 Constant *ShiftDiff = ConstantInt::get(Ty, ShrAmt - ShAmtC);
1086 auto *OldShr = cast<BinaryOperator>(Op0);
1087 auto *NewShr =
1088 BinaryOperator::Create(OldShr->getOpcode(), X, ShiftDiff);
1089 NewShr->setIsExact(OldShr->isExact());
1090 Builder.Insert(NewShr);
1092 return BinaryOperator::CreateAnd(NewShr, ConstantInt::get(Ty, Mask));
1093 }
1094 }
1095
1096 // Similar to above, but look through an intermediate trunc instruction.
1097 BinaryOperator *Shr;
1098 if (match(Op0, m_OneUse(m_Trunc(m_OneUse(m_BinOp(Shr))))) &&
1099 match(Shr, m_Shr(m_Value(X), m_APInt(C1)))) {
1100 // The larger shift direction survives through the transform.
1101 unsigned ShrAmtC = C1->getZExtValue();
1102 unsigned ShDiff = ShrAmtC > ShAmtC ? ShrAmtC - ShAmtC : ShAmtC - ShrAmtC;
1103 Constant *ShiftDiffC = ConstantInt::get(X->getType(), ShDiff);
1104 auto ShiftOpc = ShrAmtC > ShAmtC ? Shr->getOpcode() : Instruction::Shl;
1105
1106 // If C1 > C:
1107 // (trunc (X >> C1)) << C --> (trunc (X >> (C1 - C))) && (-1 << C)
1108 // If C > C1:
1109 // (trunc (X >> C1)) << C --> (trunc (X << (C - C1))) && (-1 << C)
1110 Value *NewShift = Builder.CreateBinOp(ShiftOpc, X, ShiftDiffC, "sh.diff");
1111 Value *Trunc = Builder.CreateTrunc(NewShift, Ty, "tr.sh.diff");
1113 return BinaryOperator::CreateAnd(Trunc, ConstantInt::get(Ty, Mask));
1114 }
1115
1116 if (match(Op0, m_Shl(m_Value(X), m_APInt(C1))) && C1->ult(BitWidth)) {
1117 unsigned AmtSum = ShAmtC + C1->getZExtValue();
1118 // Oversized shifts are simplified to zero in InstSimplify.
1119 if (AmtSum < BitWidth)
1120 // (X << C1) << C2 --> X << (C1 + C2)
1121 return BinaryOperator::CreateShl(X, ConstantInt::get(Ty, AmtSum));
1122 }
1123
1124 // If we have an opposite shift by the same amount, we may be able to
1125 // reorder binops and shifts to eliminate math/logic.
1126 auto isSuitableBinOpcode = [](Instruction::BinaryOps BinOpcode) {
1127 switch (BinOpcode) {
1128 default:
1129 return false;
1130 case Instruction::Add:
1131 case Instruction::And:
1132 case Instruction::Or:
1133 case Instruction::Xor:
1134 case Instruction::Sub:
1135 // NOTE: Sub is not commutable and the tranforms below may not be valid
1136 // when the shift-right is operand 1 (RHS) of the sub.
1137 return true;
1138 }
1139 };
1140 BinaryOperator *Op0BO;
1141 if (match(Op0, m_OneUse(m_BinOp(Op0BO))) &&
1142 isSuitableBinOpcode(Op0BO->getOpcode())) {
1143 // Commute so shift-right is on LHS of the binop.
1144 // (Y bop (X >> C)) << C -> ((X >> C) bop Y) << C
1145 // (Y bop ((X >> C) & CC)) << C -> (((X >> C) & CC) bop Y) << C
1146 Value *Shr = Op0BO->getOperand(0);
1147 Value *Y = Op0BO->getOperand(1);
1148 Value *X;
1149 const APInt *CC;
1150 if (Op0BO->isCommutative() && Y->hasOneUse() &&
1151 (match(Y, m_Shr(m_Value(), m_Specific(Op1))) ||
1153 m_APInt(CC)))))
1154 std::swap(Shr, Y);
1155
1156 // ((X >> C) bop Y) << C -> (X bop (Y << C)) & (~0 << C)
1157 if (match(Shr, m_OneUse(m_Shr(m_Value(X), m_Specific(Op1))))) {
1158 // Y << C
1159 Value *YS = Builder.CreateShl(Y, Op1, Op0BO->getName());
1160 // (X bop (Y << C))
1161 Value *B =
1162 Builder.CreateBinOp(Op0BO->getOpcode(), X, YS, Shr->getName());
1163 unsigned Op1Val = C->getLimitedValue(BitWidth);
1164 APInt Bits = APInt::getHighBitsSet(BitWidth, BitWidth - Op1Val);
1165 Constant *Mask = ConstantInt::get(Ty, Bits);
1166 return BinaryOperator::CreateAnd(B, Mask);
1167 }
1168
1169 // (((X >> C) & CC) bop Y) << C -> (X & (CC << C)) bop (Y << C)
1170 if (match(Shr,
1172 m_APInt(CC))))) {
1173 // Y << C
1174 Value *YS = Builder.CreateShl(Y, Op1, Op0BO->getName());
1175 // X & (CC << C)
1176 Value *M = Builder.CreateAnd(X, ConstantInt::get(Ty, CC->shl(*C)),
1177 X->getName() + ".mask");
1178 return BinaryOperator::Create(Op0BO->getOpcode(), M, YS);
1179 }
1180 }
1181
1182 // (C1 - X) << C --> (C1 << C) - (X << C)
1183 if (match(Op0, m_OneUse(m_Sub(m_APInt(C1), m_Value(X))))) {
1184 Constant *NewLHS = ConstantInt::get(Ty, C1->shl(*C));
1185 Value *NewShift = Builder.CreateShl(X, Op1);
1186 return BinaryOperator::CreateSub(NewLHS, NewShift);
1187 }
1188 }
1189
1190 if (setShiftFlags(I, Q))
1191 return &I;
1192
1193 // Transform (x >> y) << y to x & (-1 << y)
1194 // Valid for any type of right-shift.
1195 Value *X;
1196 if (match(Op0, m_OneUse(m_Shr(m_Value(X), m_Specific(Op1))))) {
1198 Value *Mask = Builder.CreateShl(AllOnes, Op1);
1199 return BinaryOperator::CreateAnd(Mask, X);
1200 }
1201
1202 Constant *C1;
1203 if (match(Op1, m_Constant(C1))) {
1204 Constant *C2;
1205 Value *X;
1206 // (X * C2) << C1 --> X * (C2 << C1)
1207 if (match(Op0, m_Mul(m_Value(X), m_Constant(C2))))
1208 return BinaryOperator::CreateMul(X, ConstantExpr::getShl(C2, C1));
1209
1210 // shl (zext i1 X), C1 --> select (X, 1 << C1, 0)
1211 if (match(Op0, m_ZExt(m_Value(X))) && X->getType()->isIntOrIntVectorTy(1)) {
1212 auto *NewC = ConstantExpr::getShl(ConstantInt::get(Ty, 1), C1);
1214 }
1215 }
1216
1217 if (match(Op0, m_One())) {
1218 // (1 << (C - x)) -> ((1 << C) >> x) if C is bitwidth - 1
1219 if (match(Op1, m_Sub(m_SpecificInt(BitWidth - 1), m_Value(X))))
1220 return BinaryOperator::CreateLShr(
1222
1223 // Canonicalize "extract lowest set bit" using cttz to and-with-negate:
1224 // 1 << (cttz X) --> -X & X
1225 if (match(Op1,
1226 m_OneUse(m_Intrinsic<Intrinsic::cttz>(m_Value(X), m_Value())))) {
1227 Value *NegX = Builder.CreateNeg(X, "neg");
1228 return BinaryOperator::CreateAnd(NegX, X);
1229 }
1230 }
1231
1232 return nullptr;
1233}
1234
1236 if (Value *V = simplifyLShrInst(I.getOperand(0), I.getOperand(1), I.isExact(),
1238 return replaceInstUsesWith(I, V);
1239
1241 return X;
1242
1244 return R;
1245
1246 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1247 Type *Ty = I.getType();
1248 Value *X;
1249 const APInt *C;
1250 unsigned BitWidth = Ty->getScalarSizeInBits();
1251
1252 // (iN (~X) u>> (N - 1)) --> zext (X > -1)
1253 if (match(Op0, m_OneUse(m_Not(m_Value(X)))) &&
1255 return new ZExtInst(Builder.CreateIsNotNeg(X, "isnotneg"), Ty);
1256
1257 if (match(Op1, m_APInt(C))) {
1258 unsigned ShAmtC = C->getZExtValue();
1259 auto *II = dyn_cast<IntrinsicInst>(Op0);
1260 if (II && isPowerOf2_32(BitWidth) && Log2_32(BitWidth) == ShAmtC &&
1261 (II->getIntrinsicID() == Intrinsic::ctlz ||
1262 II->getIntrinsicID() == Intrinsic::cttz ||
1263 II->getIntrinsicID() == Intrinsic::ctpop)) {
1264 // ctlz.i32(x)>>5 --> zext(x == 0)
1265 // cttz.i32(x)>>5 --> zext(x == 0)
1266 // ctpop.i32(x)>>5 --> zext(x == -1)
1267 bool IsPop = II->getIntrinsicID() == Intrinsic::ctpop;
1268 Constant *RHS = ConstantInt::getSigned(Ty, IsPop ? -1 : 0);
1269 Value *Cmp = Builder.CreateICmpEQ(II->getArgOperand(0), RHS);
1270 return new ZExtInst(Cmp, Ty);
1271 }
1272
1273 Value *X;
1274 const APInt *C1;
1275 if (match(Op0, m_Shl(m_Value(X), m_APInt(C1))) && C1->ult(BitWidth)) {
1276 if (C1->ult(ShAmtC)) {
1277 unsigned ShlAmtC = C1->getZExtValue();
1278 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmtC - ShlAmtC);
1279 if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) {
1280 // (X <<nuw C1) >>u C --> X >>u (C - C1)
1281 auto *NewLShr = BinaryOperator::CreateLShr(X, ShiftDiff);
1282 NewLShr->setIsExact(I.isExact());
1283 return NewLShr;
1284 }
1285 if (Op0->hasOneUse()) {
1286 // (X << C1) >>u C --> (X >>u (C - C1)) & (-1 >> C)
1287 Value *NewLShr = Builder.CreateLShr(X, ShiftDiff, "", I.isExact());
1289 return BinaryOperator::CreateAnd(NewLShr, ConstantInt::get(Ty, Mask));
1290 }
1291 } else if (C1->ugt(ShAmtC)) {
1292 unsigned ShlAmtC = C1->getZExtValue();
1293 Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmtC - ShAmtC);
1294 if (cast<BinaryOperator>(Op0)->hasNoUnsignedWrap()) {
1295 // (X <<nuw C1) >>u C --> X <<nuw/nsw (C1 - C)
1296 auto *NewShl = BinaryOperator::CreateShl(X, ShiftDiff);
1297 NewShl->setHasNoUnsignedWrap(true);
1298 NewShl->setHasNoSignedWrap(ShAmtC > 0);
1299 return NewShl;
1300 }
1301 if (Op0->hasOneUse()) {
1302 // (X << C1) >>u C --> X << (C1 - C) & (-1 >> C)
1303 Value *NewShl = Builder.CreateShl(X, ShiftDiff);
1305 return BinaryOperator::CreateAnd(NewShl, ConstantInt::get(Ty, Mask));
1306 }
1307 } else {
1308 assert(*C1 == ShAmtC);
1309 // (X << C) >>u C --> X & (-1 >>u C)
1311 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, Mask));
1312 }
1313 }
1314
1315 // ((X << C) + Y) >>u C --> (X + (Y >>u C)) & (-1 >>u C)
1316 // TODO: Consolidate with the more general transform that starts from shl
1317 // (the shifts are in the opposite order).
1318 Value *Y;
1319 if (match(Op0,
1321 m_Value(Y))))) {
1322 Value *NewLshr = Builder.CreateLShr(Y, Op1);
1323 Value *NewAdd = Builder.CreateAdd(NewLshr, X);
1324 unsigned Op1Val = C->getLimitedValue(BitWidth);
1325 APInt Bits = APInt::getLowBitsSet(BitWidth, BitWidth - Op1Val);
1326 Constant *Mask = ConstantInt::get(Ty, Bits);
1327 return BinaryOperator::CreateAnd(NewAdd, Mask);
1328 }
1329
1330 if (match(Op0, m_OneUse(m_ZExt(m_Value(X)))) &&
1331 (!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType()))) {
1332 assert(ShAmtC < X->getType()->getScalarSizeInBits() &&
1333 "Big shift not simplified to zero?");
1334 // lshr (zext iM X to iN), C --> zext (lshr X, C) to iN
1335 Value *NewLShr = Builder.CreateLShr(X, ShAmtC);
1336 return new ZExtInst(NewLShr, Ty);
1337 }
1338
1339 if (match(Op0, m_SExt(m_Value(X)))) {
1340 unsigned SrcTyBitWidth = X->getType()->getScalarSizeInBits();
1341 // lshr (sext i1 X to iN), C --> select (X, -1 >> C, 0)
1342 if (SrcTyBitWidth == 1) {
1343 auto *NewC = ConstantInt::get(
1344 Ty, APInt::getLowBitsSet(BitWidth, BitWidth - ShAmtC));
1346 }
1347
1348 if ((!Ty->isIntegerTy() || shouldChangeType(Ty, X->getType())) &&
1349 Op0->hasOneUse()) {
1350 // Are we moving the sign bit to the low bit and widening with high
1351 // zeros? lshr (sext iM X to iN), N-1 --> zext (lshr X, M-1) to iN
1352 if (ShAmtC == BitWidth - 1) {
1353 Value *NewLShr = Builder.CreateLShr(X, SrcTyBitWidth - 1);
1354 return new ZExtInst(NewLShr, Ty);
1355 }
1356
1357 // lshr (sext iM X to iN), N-M --> zext (ashr X, min(N-M, M-1)) to iN
1358 if (ShAmtC == BitWidth - SrcTyBitWidth) {
1359 // The new shift amount can't be more than the narrow source type.
1360 unsigned NewShAmt = std::min(ShAmtC, SrcTyBitWidth - 1);
1361 Value *AShr = Builder.CreateAShr(X, NewShAmt);
1362 return new ZExtInst(AShr, Ty);
1363 }
1364 }
1365 }
1366
1367 if (ShAmtC == BitWidth - 1) {
1368 // lshr i32 or(X,-X), 31 --> zext (X != 0)
1369 if (match(Op0, m_OneUse(m_c_Or(m_Neg(m_Value(X)), m_Deferred(X)))))
1370 return new ZExtInst(Builder.CreateIsNotNull(X), Ty);
1371
1372 // lshr i32 (X -nsw Y), 31 --> zext (X < Y)
1373 if (match(Op0, m_OneUse(m_NSWSub(m_Value(X), m_Value(Y)))))
1374 return new ZExtInst(Builder.CreateICmpSLT(X, Y), Ty);
1375
1376 // Check if a number is negative and odd:
1377 // lshr i32 (srem X, 2), 31 --> and (X >> 31), X
1378 if (match(Op0, m_OneUse(m_SRem(m_Value(X), m_SpecificInt(2))))) {
1379 Value *Signbit = Builder.CreateLShr(X, ShAmtC);
1380 return BinaryOperator::CreateAnd(Signbit, X);
1381 }
1382 }
1383
1384 // (X >>u C1) >>u C --> X >>u (C1 + C)
1385 if (match(Op0, m_LShr(m_Value(X), m_APInt(C1)))) {
1386 // Oversized shifts are simplified to zero in InstSimplify.
1387 unsigned AmtSum = ShAmtC + C1->getZExtValue();
1388 if (AmtSum < BitWidth)
1389 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
1390 }
1391
1392 Instruction *TruncSrc;
1393 if (match(Op0, m_OneUse(m_Trunc(m_Instruction(TruncSrc)))) &&
1394 match(TruncSrc, m_LShr(m_Value(X), m_APInt(C1)))) {
1395 unsigned SrcWidth = X->getType()->getScalarSizeInBits();
1396 unsigned AmtSum = ShAmtC + C1->getZExtValue();
1397
1398 // If the combined shift fits in the source width:
1399 // (trunc (X >>u C1)) >>u C --> and (trunc (X >>u (C1 + C)), MaskC
1400 //
1401 // If the first shift covers the number of bits truncated, then the
1402 // mask instruction is eliminated (and so the use check is relaxed).
1403 if (AmtSum < SrcWidth &&
1404 (TruncSrc->hasOneUse() || C1->uge(SrcWidth - BitWidth))) {
1405 Value *SumShift = Builder.CreateLShr(X, AmtSum, "sum.shift");
1406 Value *Trunc = Builder.CreateTrunc(SumShift, Ty, I.getName());
1407
1408 // If the first shift does not cover the number of bits truncated, then
1409 // we require a mask to get rid of high bits in the result.
1410 APInt MaskC = APInt::getAllOnes(BitWidth).lshr(ShAmtC);
1411 return BinaryOperator::CreateAnd(Trunc, ConstantInt::get(Ty, MaskC));
1412 }
1413 }
1414
1415 const APInt *MulC;
1416 if (match(Op0, m_NUWMul(m_Value(X), m_APInt(MulC)))) {
1417 // Look for a "splat" mul pattern - it replicates bits across each half of
1418 // a value, so a right shift is just a mask of the low bits:
1419 // lshr i[2N] (mul nuw X, (2^N)+1), N --> and iN X, (2^N)-1
1420 // TODO: Generalize to allow more than just half-width shifts?
1421 if (BitWidth > 2 && ShAmtC * 2 == BitWidth && (*MulC - 1).isPowerOf2() &&
1422 MulC->logBase2() == ShAmtC)
1423 return BinaryOperator::CreateAnd(X, ConstantInt::get(Ty, *MulC - 2));
1424
1425 // The one-use check is not strictly necessary, but codegen may not be
1426 // able to invert the transform and perf may suffer with an extra mul
1427 // instruction.
1428 if (Op0->hasOneUse()) {
1429 APInt NewMulC = MulC->lshr(ShAmtC);
1430 // if c is divisible by (1 << ShAmtC):
1431 // lshr (mul nuw x, MulC), ShAmtC -> mul nuw nsw x, (MulC >> ShAmtC)
1432 if (MulC->eq(NewMulC.shl(ShAmtC))) {
1433 auto *NewMul =
1434 BinaryOperator::CreateNUWMul(X, ConstantInt::get(Ty, NewMulC));
1435 assert(ShAmtC != 0 &&
1436 "lshr X, 0 should be handled by simplifyLShrInst.");
1437 NewMul->setHasNoSignedWrap(true);
1438 return NewMul;
1439 }
1440 }
1441 }
1442
1443 // Try to narrow bswap.
1444 // In the case where the shift amount equals the bitwidth difference, the
1445 // shift is eliminated.
1446 if (match(Op0, m_OneUse(m_Intrinsic<Intrinsic::bswap>(
1447 m_OneUse(m_ZExt(m_Value(X))))))) {
1448 unsigned SrcWidth = X->getType()->getScalarSizeInBits();
1449 unsigned WidthDiff = BitWidth - SrcWidth;
1450 if (SrcWidth % 16 == 0) {
1451 Value *NarrowSwap = Builder.CreateUnaryIntrinsic(Intrinsic::bswap, X);
1452 if (ShAmtC >= WidthDiff) {
1453 // (bswap (zext X)) >> C --> zext (bswap X >> C')
1454 Value *NewShift = Builder.CreateLShr(NarrowSwap, ShAmtC - WidthDiff);
1455 return new ZExtInst(NewShift, Ty);
1456 } else {
1457 // (bswap (zext X)) >> C --> (zext (bswap X)) << C'
1458 Value *NewZExt = Builder.CreateZExt(NarrowSwap, Ty);
1459 Constant *ShiftDiff = ConstantInt::get(Ty, WidthDiff - ShAmtC);
1460 return BinaryOperator::CreateShl(NewZExt, ShiftDiff);
1461 }
1462 }
1463 }
1464
1465 // Reduce add-carry of bools to logic:
1466 // ((zext BoolX) + (zext BoolY)) >> 1 --> zext (BoolX && BoolY)
1467 Value *BoolX, *BoolY;
1468 if (ShAmtC == 1 && match(Op0, m_Add(m_Value(X), m_Value(Y))) &&
1469 match(X, m_ZExt(m_Value(BoolX))) && match(Y, m_ZExt(m_Value(BoolY))) &&
1470 BoolX->getType()->isIntOrIntVectorTy(1) &&
1471 BoolY->getType()->isIntOrIntVectorTy(1) &&
1472 (X->hasOneUse() || Y->hasOneUse() || Op0->hasOneUse())) {
1473 Value *And = Builder.CreateAnd(BoolX, BoolY);
1474 return new ZExtInst(And, Ty);
1475 }
1476 }
1477
1479 if (setShiftFlags(I, Q))
1480 return &I;
1481
1482 // Transform (x << y) >> y to x & (-1 >> y)
1483 if (match(Op0, m_OneUse(m_Shl(m_Value(X), m_Specific(Op1))))) {
1485 Value *Mask = Builder.CreateLShr(AllOnes, Op1);
1486 return BinaryOperator::CreateAnd(Mask, X);
1487 }
1488
1489 if (Instruction *Overflow = foldLShrOverflowBit(I))
1490 return Overflow;
1491
1492 return nullptr;
1493}
1494
1497 BinaryOperator &OldAShr) {
1498 assert(OldAShr.getOpcode() == Instruction::AShr &&
1499 "Must be called with arithmetic right-shift instruction only.");
1500
1501 // Check that constant C is a splat of the element-wise bitwidth of V.
1502 auto BitWidthSplat = [](Constant *C, Value *V) {
1503 return match(
1505 APInt(C->getType()->getScalarSizeInBits(),
1506 V->getType()->getScalarSizeInBits())));
1507 };
1508
1509 // It should look like variable-length sign-extension on the outside:
1510 // (Val << (bitwidth(Val)-Nbits)) a>> (bitwidth(Val)-Nbits)
1511 Value *NBits;
1512 Instruction *MaybeTrunc;
1513 Constant *C1, *C2;
1514 if (!match(&OldAShr,
1515 m_AShr(m_Shl(m_Instruction(MaybeTrunc),
1517 m_ZExtOrSelf(m_Value(NBits))))),
1519 m_ZExtOrSelf(m_Deferred(NBits)))))) ||
1520 !BitWidthSplat(C1, &OldAShr) || !BitWidthSplat(C2, &OldAShr))
1521 return nullptr;
1522
1523 // There may or may not be a truncation after outer two shifts.
1524 Instruction *HighBitExtract;
1525 match(MaybeTrunc, m_TruncOrSelf(m_Instruction(HighBitExtract)));
1526 bool HadTrunc = MaybeTrunc != HighBitExtract;
1527
1528 // And finally, the innermost part of the pattern must be a right-shift.
1529 Value *X, *NumLowBitsToSkip;
1530 if (!match(HighBitExtract, m_Shr(m_Value(X), m_Value(NumLowBitsToSkip))))
1531 return nullptr;
1532
1533 // Said right-shift must extract high NBits bits - C0 must be it's bitwidth.
1534 Constant *C0;
1535 if (!match(NumLowBitsToSkip,
1537 m_Sub(m_Constant(C0), m_ZExtOrSelf(m_Specific(NBits))))) ||
1538 !BitWidthSplat(C0, HighBitExtract))
1539 return nullptr;
1540
1541 // Since the NBits is identical for all shifts, if the outermost and
1542 // innermost shifts are identical, then outermost shifts are redundant.
1543 // If we had truncation, do keep it though.
1544 if (HighBitExtract->getOpcode() == OldAShr.getOpcode())
1545 return replaceInstUsesWith(OldAShr, MaybeTrunc);
1546
1547 // Else, if there was a truncation, then we need to ensure that one
1548 // instruction will go away.
1549 if (HadTrunc && !match(&OldAShr, m_c_BinOp(m_OneUse(m_Value()), m_Value())))
1550 return nullptr;
1551
1552 // Finally, bypass two innermost shifts, and perform the outermost shift on
1553 // the operands of the innermost shift.
1554 Instruction *NewAShr =
1555 BinaryOperator::Create(OldAShr.getOpcode(), X, NumLowBitsToSkip);
1556 NewAShr->copyIRFlags(HighBitExtract); // We can preserve 'exact'-ness.
1557 if (!HadTrunc)
1558 return NewAShr;
1559
1560 Builder.Insert(NewAShr);
1561 return TruncInst::CreateTruncOrBitCast(NewAShr, OldAShr.getType());
1562}
1563
1565 if (Value *V = simplifyAShrInst(I.getOperand(0), I.getOperand(1), I.isExact(),
1567 return replaceInstUsesWith(I, V);
1568
1570 return X;
1571
1573 return R;
1574
1575 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1576 Type *Ty = I.getType();
1577 unsigned BitWidth = Ty->getScalarSizeInBits();
1578 const APInt *ShAmtAPInt;
1579 if (match(Op1, m_APInt(ShAmtAPInt)) && ShAmtAPInt->ult(BitWidth)) {
1580 unsigned ShAmt = ShAmtAPInt->getZExtValue();
1581
1582 // If the shift amount equals the difference in width of the destination
1583 // and source scalar types:
1584 // ashr (shl (zext X), C), C --> sext X
1585 Value *X;
1586 if (match(Op0, m_Shl(m_ZExt(m_Value(X)), m_Specific(Op1))) &&
1587 ShAmt == BitWidth - X->getType()->getScalarSizeInBits())
1588 return new SExtInst(X, Ty);
1589
1590 // We can't handle (X << C1) >>s C2. It shifts arbitrary bits in. However,
1591 // we can handle (X <<nsw C1) >>s C2 since it only shifts in sign bits.
1592 const APInt *ShOp1;
1593 if (match(Op0, m_NSWShl(m_Value(X), m_APInt(ShOp1))) &&
1594 ShOp1->ult(BitWidth)) {
1595 unsigned ShlAmt = ShOp1->getZExtValue();
1596 if (ShlAmt < ShAmt) {
1597 // (X <<nsw C1) >>s C2 --> X >>s (C2 - C1)
1598 Constant *ShiftDiff = ConstantInt::get(Ty, ShAmt - ShlAmt);
1599 auto *NewAShr = BinaryOperator::CreateAShr(X, ShiftDiff);
1600 NewAShr->setIsExact(I.isExact());
1601 return NewAShr;
1602 }
1603 if (ShlAmt > ShAmt) {
1604 // (X <<nsw C1) >>s C2 --> X <<nsw (C1 - C2)
1605 Constant *ShiftDiff = ConstantInt::get(Ty, ShlAmt - ShAmt);
1606 auto *NewShl = BinaryOperator::Create(Instruction::Shl, X, ShiftDiff);
1607 NewShl->setHasNoSignedWrap(true);
1608 return NewShl;
1609 }
1610 }
1611
1612 if (match(Op0, m_AShr(m_Value(X), m_APInt(ShOp1))) &&
1613 ShOp1->ult(BitWidth)) {
1614 unsigned AmtSum = ShAmt + ShOp1->getZExtValue();
1615 // Oversized arithmetic shifts replicate the sign bit.
1616 AmtSum = std::min(AmtSum, BitWidth - 1);
1617 // (X >>s C1) >>s C2 --> X >>s (C1 + C2)
1618 return BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
1619 }
1620
1621 if (match(Op0, m_OneUse(m_SExt(m_Value(X)))) &&
1622 (Ty->isVectorTy() || shouldChangeType(Ty, X->getType()))) {
1623 // ashr (sext X), C --> sext (ashr X, C')
1624 Type *SrcTy = X->getType();
1625 ShAmt = std::min(ShAmt, SrcTy->getScalarSizeInBits() - 1);
1626 Value *NewSh = Builder.CreateAShr(X, ConstantInt::get(SrcTy, ShAmt));
1627 return new SExtInst(NewSh, Ty);
1628 }
1629
1630 if (ShAmt == BitWidth - 1) {
1631 // ashr i32 or(X,-X), 31 --> sext (X != 0)
1632 if (match(Op0, m_OneUse(m_c_Or(m_Neg(m_Value(X)), m_Deferred(X)))))
1633 return new SExtInst(Builder.CreateIsNotNull(X), Ty);
1634
1635 // ashr i32 (X -nsw Y), 31 --> sext (X < Y)
1636 Value *Y;
1637 if (match(Op0, m_OneUse(m_NSWSub(m_Value(X), m_Value(Y)))))
1638 return new SExtInst(Builder.CreateICmpSLT(X, Y), Ty);
1639 }
1640 }
1641
1643 if (setShiftFlags(I, Q))
1644 return &I;
1645
1646 // Prefer `-(x & 1)` over `(x << (bitwidth(x)-1)) a>> (bitwidth(x)-1)`
1647 // as the pattern to splat the lowest bit.
1648 // FIXME: iff X is already masked, we don't need the one-use check.
1649 Value *X;
1650 if (match(Op1, m_SpecificIntAllowUndef(BitWidth - 1)) &&
1653 Constant *Mask = ConstantInt::get(Ty, 1);
1654 // Retain the knowledge about the ignored lanes.
1656 Constant::mergeUndefsWith(Mask, cast<Constant>(Op1)),
1657 cast<Constant>(cast<Instruction>(Op0)->getOperand(1)));
1658 X = Builder.CreateAnd(X, Mask);
1660 }
1661
1663 return R;
1664
1665 // See if we can turn a signed shr into an unsigned shr.
1667 Instruction *Lshr = BinaryOperator::CreateLShr(Op0, Op1);
1668 Lshr->setIsExact(I.isExact());
1669 return Lshr;
1670 }
1671
1672 // ashr (xor %x, -1), %y --> xor (ashr %x, %y), -1
1673 if (match(Op0, m_OneUse(m_Not(m_Value(X))))) {
1674 // Note that we must drop 'exact'-ness of the shift!
1675 // Note that we can't keep undef's in -1 vector constant!
1676 auto *NewAShr = Builder.CreateAShr(X, Op1, Op0->getName() + ".not");
1677 return BinaryOperator::CreateNot(NewAShr);
1678 }
1679
1680 return nullptr;
1681}
MachineBasicBlock MachineBasicBlock::iterator DebugLoc DL
static GCRegistry::Add< OcamlGC > B("ocaml", "ocaml 3.10-compatible GC")
static GCRegistry::Add< ErlangGC > A("erlang", "erlang-compatible garbage collector")
#define LLVM_DEBUG(X)
Definition: Debug.h:101
static GCMetadataPrinterRegistry::Add< ErlangGCPrinter > X("erlang", "erlang-compatible garbage collector")
This file provides internal interfaces used to implement the InstCombine.
static Value * foldShiftedShift(BinaryOperator *InnerShift, unsigned OuterShAmt, bool IsOuterShl, InstCombiner::BuilderTy &Builder)
Fold OuterShift (InnerShift X, C1), C2.
static bool setShiftFlags(BinaryOperator &I, const SimplifyQuery &Q)
static Instruction * dropRedundantMaskingOfLeftShiftInput(BinaryOperator *OuterShift, const SimplifyQuery &Q, InstCombiner::BuilderTy &Builder)
static bool canEvaluateShifted(Value *V, unsigned NumBits, bool IsLeftShift, InstCombinerImpl &IC, Instruction *CxtI)
See if we can compute the specified value, but shifted logically to the left or right by some number ...
bool canTryToConstantAddTwoShiftAmounts(Value *Sh0, Value *ShAmt0, Value *Sh1, Value *ShAmt1)
static Instruction * foldShiftOfShiftedBinOp(BinaryOperator &I, InstCombiner::BuilderTy &Builder)
If we have a shift-by-constant of a bin op (bitwise logic op or add/sub w/ shl) that itself has a shi...
static bool canEvaluateShiftedShift(unsigned OuterShAmt, bool IsOuterShl, Instruction *InnerShift, InstCombinerImpl &IC, Instruction *CxtI)
Return true if we can simplify two logical (either left or right) shifts that have constant shift amo...
static Value * getShiftedValue(Value *V, unsigned NumBits, bool isLeftShift, InstCombinerImpl &IC, const DataLayout &DL)
When canEvaluateShifted() returns true for an expression, this function inserts the new computation t...
static bool canShiftBinOpWithConstantRHS(BinaryOperator &Shift, BinaryOperator *BO)
This file provides the interface for the instcombine pass implementation.
static bool hasNoUnsignedWrap(BinaryOperator &I)
#define I(x, y, z)
Definition: MD5.cpp:58
static GCMetadataPrinterRegistry::Add< OcamlGCMetadataPrinter > Y("ocaml", "ocaml 3.10-compatible collector")
const SmallVectorImpl< MachineOperand > & Cond
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
static unsigned getScalarSizeInBits(Type *Ty)
static SymbolRef::Type getType(const Symbol *Sym)
Definition: TapiFile.cpp:40
static std::optional< unsigned > getOpcode(ArrayRef< VPValue * > Values)
Returns the opcode of Values or ~0 if they do not all agree.
Definition: VPlanSLP.cpp:191
Value * RHS
Class for arbitrary precision integers.
Definition: APInt.h:76
static APInt getAllOnes(unsigned numBits)
Return an APInt of a specified width with all bits set.
Definition: APInt.h:212
bool isNegatedPowerOf2() const
Check if this APInt's negated value is a power of two greater than zero.
Definition: APInt.h:427
static APInt getSignMask(unsigned BitWidth)
Get the SignMask for a specific bit width.
Definition: APInt.h:207
bool isMinSignedValue() const
Determine if this is the smallest signed value.
Definition: APInt.h:401
uint64_t getZExtValue() const
Get zero extended value.
Definition: APInt.h:1485
bool ugt(const APInt &RHS) const
Unsigned greater than comparison.
Definition: APInt.h:1154
bool isZero() const
Determine if this value is zero, i.e. all bits are clear.
Definition: APInt.h:358
bool ult(const APInt &RHS) const
Unsigned less than comparison.
Definition: APInt.h:1083
bool isNegative() const
Determine sign of this APInt.
Definition: APInt.h:307
bool eq(const APInt &RHS) const
Equality comparison.
Definition: APInt.h:1051
unsigned countr_zero() const
Count the number of trailing zero bits.
Definition: APInt.h:1583
unsigned logBase2() const
Definition: APInt.h:1696
uint64_t getLimitedValue(uint64_t Limit=UINT64_MAX) const
If this value is smaller than the specified limit, return it, otherwise return the limit value.
Definition: APInt.h:453
APInt shl(unsigned shiftAmt) const
Left-shift function.
Definition: APInt.h:851
static APInt getLowBitsSet(unsigned numBits, unsigned loBitsSet)
Constructs an APInt value that has the bottom loBitsSet bits set.
Definition: APInt.h:284
static APInt getHighBitsSet(unsigned numBits, unsigned hiBitsSet)
Constructs an APInt value that has the top hiBitsSet bits set.
Definition: APInt.h:274
APInt lshr(unsigned shiftAmt) const
Logical right-shift function.
Definition: APInt.h:829
bool uge(const APInt &RHS) const
Unsigned greater or equal comparison.
Definition: APInt.h:1193
static BinaryOperator * Create(BinaryOps Op, Value *S1, Value *S2, const Twine &Name=Twine(), Instruction *InsertBefore=nullptr)
Construct a binary instruction, given the opcode and the two operands.
static BinaryOperator * CreateNeg(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
Helper functions to construct and inspect unary operations (NEG and NOT) via binary operators SUB and...
BinaryOps getOpcode() const
Definition: InstrTypes.h:402
static BinaryOperator * CreateNot(Value *Op, const Twine &Name="", Instruction *InsertBefore=nullptr)
static CastInst * Create(Instruction::CastOps, Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Provides a way to construct any of the CastInst subclasses using an opcode instead of the subclass's ...
static CastInst * CreateTruncOrBitCast(Value *S, Type *Ty, const Twine &Name="", Instruction *InsertBefore=nullptr)
Create a Trunc or BitCast cast instruction.
Predicate
This enumeration lists the possible predicates for CmpInst subclasses.
Definition: InstrTypes.h:780
@ ICMP_SLE
signed less or equal
Definition: InstrTypes.h:810
@ ICMP_ULT
unsigned less than
Definition: InstrTypes.h:805
@ ICMP_EQ
equal
Definition: InstrTypes.h:801
@ ICMP_SGE
signed greater or equal
Definition: InstrTypes.h:808
static Constant * getSub(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2460
static Constant * getNot(Constant *C)
Definition: Constants.cpp:2447
static Constant * getShl(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2478
static Constant * getAdd(Constant *C1, Constant *C2, bool HasNUW=false, bool HasNSW=false)
Definition: Constants.cpp:2453
static Constant * getTrunc(Constant *C, Type *Ty, bool OnlyIfReduced=false)
Definition: Constants.cpp:2014
static Constant * get(Type *Ty, uint64_t V, bool IsSigned=false)
If Ty is a vector type, return a Constant with a splat of the given value.
Definition: Constants.cpp:888
static ConstantInt * getSigned(IntegerType *Ty, int64_t V)
Return a ConstantInt with the specified value for the specified type.
Definition: Constants.h:115
This is an important base class in LLVM.
Definition: Constant.h:41
static Constant * replaceUndefsWith(Constant *C, Constant *Replacement)
Try to replace undefined constant C or undefined elements in C with Replacement.
Definition: Constants.cpp:753
static Constant * mergeUndefsWith(Constant *C, Constant *Other)
Merges undefs of a Constant with another Constant, along with the undefs already present.
Definition: Constants.cpp:777
static Constant * getAllOnesValue(Type *Ty)
Definition: Constants.cpp:403
static Constant * getNullValue(Type *Ty)
Constructor to create a '0' constant of arbitrary type.
Definition: Constants.cpp:356
A parsed version of the target data layout string in and methods for querying it.
Definition: DataLayout.h:110
CallInst * CreateUnaryIntrinsic(Intrinsic::ID ID, Value *V, Instruction *FMFSource=nullptr, const Twine &Name="")
Create a call to intrinsic ID with 1 operand which is mangled on its type.
Definition: IRBuilder.cpp:913
Value * CreateICmpULT(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:2230
Value * CreateTrunc(Value *V, Type *DestTy, const Twine &Name="")
Definition: IRBuilder.h:1996
Value * CreateNeg(Value *V, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1715
Value * CreateLShr(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Definition: IRBuilder.h:1431
Value * CreateIsNotNeg(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg > -1.
Definition: IRBuilder.h:2532
Value * CreateICmpEQ(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:2214
InstTy * Insert(InstTy *I, const Twine &Name="") const
Insert and return the specified instruction.
Definition: IRBuilder.h:145
Value * CreateShl(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1410
Value * CreateZExt(Value *V, Type *DestTy, const Twine &Name="", bool IsNonNeg=false)
Definition: IRBuilder.h:2000
Value * CreateAnd(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:1469
Value * CreateAdd(Value *LHS, Value *RHS, const Twine &Name="", bool HasNUW=false, bool HasNSW=false)
Definition: IRBuilder.h:1321
Value * CreateIsNotNull(Value *Arg, const Twine &Name="")
Return a boolean value testing if Arg != 0.
Definition: IRBuilder.h:2522
Value * CreateBinOp(Instruction::BinaryOps Opc, Value *LHS, Value *RHS, const Twine &Name="", MDNode *FPMathTag=nullptr)
Definition: IRBuilder.h:1660
Value * CreateICmpSLT(Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:2246
void SetInsertPoint(BasicBlock *TheBB)
This specifies that created instructions should be appended to the end of the specified block.
Definition: IRBuilder.h:180
Value * CreateAShr(Value *LHS, Value *RHS, const Twine &Name="", bool isExact=false)
Definition: IRBuilder.h:1450
Value * CreateICmp(CmpInst::Predicate P, Value *LHS, Value *RHS, const Twine &Name="")
Definition: IRBuilder.h:2324
Instruction * FoldOpIntoSelect(Instruction &Op, SelectInst *SI, bool FoldWithMultiUse=false)
Given an instruction with a select as one operand and a constant as the other operand,...
Instruction * visitLShr(BinaryOperator &I)
Instruction * foldBinOpIntoSelectOrPhi(BinaryOperator &I)
This is a convenience wrapper function for the above two functions.
Value * reassociateShiftAmtsOfTwoSameDirectionShifts(BinaryOperator *Sh0, const SimplifyQuery &SQ, bool AnalyzeForSignBitExtraction=false)
Instruction * visitAShr(BinaryOperator &I)
Instruction * eraseInstFromFunction(Instruction &I) override
Combiner aware instruction erasure.
Instruction * visitShl(BinaryOperator &I)
Instruction * foldBinopWithPhiOperands(BinaryOperator &BO)
For a binary operator with 2 phi operands, try to hoist the binary operation before the phi.
Instruction * foldVariableSignZeroExtensionOfVariableHighBitExtract(BinaryOperator &OldAShr)
Instruction * commonShiftTransforms(BinaryOperator &I)
bool SimplifyDemandedInstructionBits(Instruction &Inst)
Tries to simplify operands to an integer instruction based on its demanded bits.
Instruction * foldVectorBinop(BinaryOperator &Inst)
Canonicalize the position of binops relative to shufflevector.
Instruction * FoldShiftByConstant(Value *Op0, Constant *Op1, BinaryOperator &I)
SimplifyQuery SQ
Definition: InstCombiner.h:76
Instruction * replaceInstUsesWith(Instruction &I, Value *V)
A combiner-aware RAUW-like routine.
Definition: InstCombiner.h:424
Instruction * InsertNewInstWith(Instruction *New, BasicBlock::iterator Old)
Same as InsertNewInstBefore, but also sets the debug loc.
Definition: InstCombiner.h:413
const DataLayout & DL
Definition: InstCombiner.h:75
AssumptionCache & AC
Definition: InstCombiner.h:72
void addToWorklist(Instruction *I)
Definition: InstCombiner.h:374
Instruction * replaceOperand(Instruction &I, unsigned OpNum, Value *V)
Replace operand of instruction and add old operand to the worklist.
Definition: InstCombiner.h:448
BuilderTy & Builder
Definition: InstCombiner.h:60
bool MaskedValueIsZero(const Value *V, const APInt &Mask, unsigned Depth=0, const Instruction *CxtI=nullptr) const
Definition: InstCombiner.h:485
void setHasNoUnsignedWrap(bool b=true)
Set or clear the nuw flag on this instruction, which must be an operator which supports this flag.
bool hasNoUnsignedWrap() const LLVM_READONLY
Determine whether the no unsigned wrap flag is set.
bool hasNoSignedWrap() const LLVM_READONLY
Determine whether the no signed wrap flag is set.
void copyIRFlags(const Value *V, bool IncludeWrapFlags=true)
Convenience method to copy supported exact, fast-math, and (optionally) wrapping flags from V to this...
void setHasNoSignedWrap(bool b=true)
Set or clear the nsw flag on this instruction, which must be an operator which supports this flag.
bool isCommutative() const LLVM_READONLY
Return true if the instruction is commutative:
bool isExact() const LLVM_READONLY
Determine whether the exact flag is set.
bool isLogicalShift() const
Return true if this is a logical shift left or a logical shift right.
Definition: Instruction.h:278
unsigned getOpcode() const
Returns a member of one of the enums like Instruction::Add.
Definition: Instruction.h:239
void setIsExact(bool b=true)
Set or clear the exact flag on this instruction, which must be an operator which supports this flag.
op_range incoming_values()
void setIncomingValue(unsigned i, Value *V)
Value * getIncomingValue(unsigned i) const
Return incoming value number x.
unsigned getNumIncomingValues() const
Return the number of incoming edges.
This class represents a sign extension of integer types.
This class represents the LLVM 'select' instruction.
static SelectInst * Create(Value *C, Value *S1, Value *S2, const Twine &NameStr="", Instruction *InsertBefore=nullptr, Instruction *MDFrom=nullptr)
This class represents a truncation of integer types.
The instances of the Type class are immutable: once they are created, they are never changed.
Definition: Type.h:45
bool isVectorTy() const
True if this is an instance of VectorType.
Definition: Type.h:265
bool isIntOrIntVectorTy() const
Return true if this is an integer type or a vector of integer types.
Definition: Type.h:234
unsigned getScalarSizeInBits() const LLVM_READONLY
If this is a vector type, return the getPrimitiveSizeInBits value for the element type.
Type * getExtendedType() const
Given scalar/vector integer type, returns a type with elements twice as wide as in the original type.
bool isIntegerTy() const
True if this is an instance of IntegerType.
Definition: Type.h:228
void setOperand(unsigned i, Value *Val)
Definition: User.h:174
Value * getOperand(unsigned i) const
Definition: User.h:169
LLVM Value Representation.
Definition: Value.h:74
Type * getType() const
All values are typed, get the type of this value.
Definition: Value.h:255
bool hasOneUse() const
Return true if there is exactly one use of this value.
Definition: Value.h:434
StringRef getName() const
Return a constant reference to the value's name.
Definition: Value.cpp:309
void takeName(Value *V)
Transfer the name from V to this value.
Definition: Value.cpp:383
This class represents zero extension of integer types.
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
@ C
The default llvm calling convention, compatible with C.
Definition: CallingConv.h:34
cst_pred_ty< is_all_ones > m_AllOnes()
Match an integer or vector with all bits set.
Definition: PatternMatch.h:461
BinaryOp_match< LHS, RHS, Instruction::And > m_And(const LHS &L, const RHS &R)
specific_intval< false > m_SpecificInt(APInt V)
Match a specific integer value or vector with all elements equal to the value.
Definition: PatternMatch.h:862
BinaryOp_match< LHS, RHS, Instruction::Add > m_Add(const LHS &L, const RHS &R)
Definition: PatternMatch.h:982
class_match< BinaryOperator > m_BinOp()
Match an arbitrary binary operation and ignore it.
Definition: PatternMatch.h:84
OverflowingBinaryOp_match< LHS, RHS, Instruction::Add, OverflowingBinaryOperator::NoUnsignedWrap > m_NUWAdd(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::AShr > m_AShr(const LHS &L, const RHS &R)
cst_pred_ty< is_power2 > m_Power2()
Match an integer or vector power-of-2.
Definition: PatternMatch.h:552
class_match< Constant > m_Constant()
Match an arbitrary Constant and ignore it.
Definition: PatternMatch.h:144
BinaryOp_match< LHS, RHS, Instruction::And, true > m_c_And(const LHS &L, const RHS &R)
Matches an And with LHS and RHS in either order.
BinaryOp_match< LHS, RHS, Instruction::Xor > m_Xor(const LHS &L, const RHS &R)
OverflowingBinaryOp_match< LHS, RHS, Instruction::Sub, OverflowingBinaryOperator::NoSignedWrap > m_NSWSub(const LHS &L, const RHS &R)
bool match(Val *V, const Pattern &P)
Definition: PatternMatch.h:49
bind_ty< Instruction > m_Instruction(Instruction *&I)
Match an instruction, capturing it if we match.
Definition: PatternMatch.h:724
specificval_ty m_Specific(const Value *V)
Match if we have a specific specified value.
Definition: PatternMatch.h:780
BinOpPred_match< LHS, RHS, is_right_shift_op > m_Shr(const LHS &L, const RHS &R)
Matches logical shift operations.
cst_pred_ty< is_one > m_One()
Match an integer 1 or a vector with all elements equal to 1.
Definition: PatternMatch.h:525
ThreeOps_match< Cond, LHS, RHS, Instruction::Select > m_Select(const Cond &C, const LHS &L, const RHS &R)
Matches SelectInst.
match_combine_and< LTy, RTy > m_CombineAnd(const LTy &L, const RTy &R)
Combine two pattern matchers matching L && R.
Definition: PatternMatch.h:224
CastInst_match< OpTy, Instruction::ZExt > m_ZExt(const OpTy &Op)
Matches ZExt.
CastOperator_match< OpTy, Instruction::Trunc > m_Trunc(const OpTy &Op)
Matches Trunc.
BinaryOp_match< LHS, RHS, Instruction::Mul > m_Mul(const LHS &L, const RHS &R)
deferredval_ty< Value > m_Deferred(Value *const &V)
Like m_Specific(), but works if the specific value to match is determined as part of the same match()...
Definition: PatternMatch.h:798
CastInst_match< OpTy, Instruction::SExt > m_SExt(const OpTy &Op)
Matches SExt.
OneUse_match< T > m_OneUse(const T &SubPattern)
Definition: PatternMatch.h:67
BinaryOp_match< cst_pred_ty< is_zero_int >, ValTy, Instruction::Sub > m_Neg(const ValTy &V)
Matches a 'Neg' as 'sub 0, V'.
match_combine_and< class_match< Constant >, match_unless< constantexpr_match > > m_ImmConstant()
Match an arbitrary immediate Constant and ignore it.
Definition: PatternMatch.h:759
OverflowingBinaryOp_match< LHS, RHS, Instruction::Shl, OverflowingBinaryOperator::NoSignedWrap > m_NSWShl(const LHS &L, const RHS &R)
OverflowingBinaryOp_match< LHS, RHS, Instruction::Mul, OverflowingBinaryOperator::NoUnsignedWrap > m_NUWMul(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::Add, true > m_c_Add(const LHS &L, const RHS &R)
Matches a Add with LHS and RHS in either order.
match_combine_or< CastOperator_match< OpTy, Instruction::Trunc >, OpTy > m_TruncOrSelf(const OpTy &Op)
BinaryOp_match< LHS, RHS, Instruction::SDiv > m_SDiv(const LHS &L, const RHS &R)
specific_intval< true > m_SpecificIntAllowUndef(APInt V)
Definition: PatternMatch.h:870
apint_match m_APInt(const APInt *&Res)
Match a ConstantInt or splatted ConstantVector, binding the specified pointer to the contained APInt.
Definition: PatternMatch.h:278
class_match< Value > m_Value()
Match an arbitrary value and ignore it.
Definition: PatternMatch.h:76
AnyBinaryOp_match< LHS, RHS, true > m_c_BinOp(const LHS &L, const RHS &R)
Matches a BinaryOperator with LHS and RHS in either order.
BinaryOp_match< LHS, RHS, Instruction::LShr > m_LShr(const LHS &L, const RHS &R)
Exact_match< T > m_Exact(const T &SubPattern)
BinOpPred_match< LHS, RHS, is_shift_op > m_Shift(const LHS &L, const RHS &R)
Matches shift operations.
BinaryOp_match< LHS, RHS, Instruction::Shl > m_Shl(const LHS &L, const RHS &R)
BinaryOp_match< LHS, RHS, Instruction::SRem > m_SRem(const LHS &L, const RHS &R)
BinaryOp_match< cst_pred_ty< is_all_ones >, ValTy, Instruction::Xor, true > m_Not(const ValTy &V)
Matches a 'Not' as 'xor V, -1' or 'xor -1, V'.
BinaryOp_match< LHS, RHS, Instruction::Or > m_Or(const LHS &L, const RHS &R)
match_combine_or< CastInst_match< OpTy, Instruction::ZExt >, OpTy > m_ZExtOrSelf(const OpTy &Op)
BinaryOp_match< LHS, RHS, Instruction::Or, true > m_c_Or(const LHS &L, const RHS &R)
Matches an Or with LHS and RHS in either order.
BinaryOp_match< LHS, RHS, Instruction::Sub > m_Sub(const LHS &L, const RHS &R)
Definition: PatternMatch.h:994
match_combine_or< LTy, RTy > m_CombineOr(const LTy &L, const RTy &R)
Combine two pattern matchers matching L || R.
Definition: PatternMatch.h:218
cst_pred_ty< icmp_pred_with_threshold > m_SpecificInt_ICMP(ICmpInst::Predicate Predicate, const APInt &Threshold)
Match an integer or vector with every element comparing 'pred' (eg/ne/...) to Threshold.
Definition: PatternMatch.h:606
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
Value * simplifyAShrInst(Value *Op0, Value *Op1, bool IsExact, const SimplifyQuery &Q)
Given operands for a AShr, fold the result or return nulll.
Value * simplifySubInst(Value *LHS, Value *RHS, bool IsNSW, bool IsNUW, const SimplifyQuery &Q)
Given operands for a Sub, fold the result or return null.
Value * simplifyAddInst(Value *LHS, Value *RHS, bool IsNSW, bool IsNUW, const SimplifyQuery &Q)
Given operands for an Add, fold the result or return null.
unsigned Log2_32(uint32_t Value)
Return the floor log base 2 of the specified value, -1 if the value is zero.
Definition: MathExtras.h:313
Value * simplifyShlInst(Value *Op0, Value *Op1, bool IsNSW, bool IsNUW, const SimplifyQuery &Q)
Given operands for a Shl, fold the result or return null.
constexpr bool isPowerOf2_32(uint32_t Value)
Return true if the argument is a power of two > 0.
Definition: MathExtras.h:264
Value * simplifyLShrInst(Value *Op0, Value *Op1, bool IsExact, const SimplifyQuery &Q)
Given operands for a LShr, fold the result or return null.
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:163
Constant * ConstantFoldCastOperand(unsigned Opcode, Constant *C, Type *DestTy, const DataLayout &DL)
Attempt to constant fold a cast with the specified operand.
Constant * ConstantFoldBinaryOpOperands(unsigned Opcode, Constant *LHS, Constant *RHS, const DataLayout &DL)
Attempt to constant fold a binary operation with the specified operands.
@ And
Bitwise or logical AND of integers.
@ Add
Sum of integers.
void computeKnownBits(const Value *V, KnownBits &Known, const DataLayout &DL, unsigned Depth=0, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, bool UseInstrInfo=true)
Determine which bits of V are known to be either zero or one and return them in the KnownZero/KnownOn...
constexpr unsigned BitWidth
Definition: BitmaskEnum.h:191
unsigned ComputeNumSignBits(const Value *Op, const DataLayout &DL, unsigned Depth=0, AssumptionCache *AC=nullptr, const Instruction *CxtI=nullptr, const DominatorTree *DT=nullptr, bool UseInstrInfo=true)
Return the number of times the sign bit of the register is replicated into the other bits.
void swap(llvm::BitVector &LHS, llvm::BitVector &RHS)
Implement std::swap in terms of BitVector swap.
Definition: BitVector.h:860
unsigned countMinSignBits() const
Returns the number of times the sign bit is replicated into the other bits.
Definition: KnownBits.h:246
unsigned countMinTrailingZeros() const
Returns the minimum number of trailing zero bits.
Definition: KnownBits.h:233
unsigned getBitWidth() const
Get the bit width of this value.
Definition: KnownBits.h:40
unsigned countMinLeadingZeros() const
Returns the minimum number of leading zero bits.
Definition: KnownBits.h:239
APInt getMaxValue() const
Return the maximal unsigned value possible given these KnownBits.
Definition: KnownBits.h:136
const DataLayout & DL
Definition: SimplifyQuery.h:61
const Instruction * CxtI
Definition: SimplifyQuery.h:65
const DominatorTree * DT
Definition: SimplifyQuery.h:63
SimplifyQuery getWithInstruction(const Instruction *I) const
Definition: SimplifyQuery.h:96
AssumptionCache * AC
Definition: SimplifyQuery.h:64